21PM_R1_14 30_Ronald Bradford_Improving Performance With Better Indexes

合集下载

Buenos Aires. Argentina.

______This work was partially supported by the UBACYT research project TX-004, “Concurrency in distributed systems”.CELL-DEVS WITH EXPLICIT DELAYS: PARALLEL SIMULATIONGabriel A. Wainer Daniel Rodríguez Norbert Giambiasi {gabrielw, drodrigu}@dc.uba.ar Norbert.Giambiasi@ iuspim.u-3mrs.frDepartamento de Computación Facultad de Ciencias Exactas y Naturales Universidad de Buenos Aires (1428) Pabellón I. Ciudad Universitaria. Buenos Aires. Argentina.DIAM-IUSPIM Av. Escadrille Normandie Niemen13397 Marseille Cedex 20 FranceAbstractThis work describes some of the extensions included in a tool to study, model and simulate cellular models. The environment is based on the DEVS and Cell-DEVS paradigms. Cell based systems can be built using a specification language, allowing reductions in the development, checking and maintenance times of the components. A mapping between the simulation mechanism and asynchronous PDES techniques is also presented, permitting to outline the implementation of executable models in parallel.1. INTRODUCTIONIn [Zei76], the DEVS (Discrete EVents Systems specifications) formalism was proposed to model discrete events systems. A DEVS model is built using a set of behavioral models called Atomic , which can be combined to form Coupled ones. Cell-DEVS [WG98] is a paradigm that has extended the DEVS formalism, allowing the implementation of cellular models. In this formalism, each cell is defined as an atomic model using transport or inertial delays [Gho96] [GIA76]. A coupled model that includes a group of these cells will form a cellular model. CD++ [BBW98] is a tool that allows implementing the theoretical concepts specified by the DEVS and Cell-DEVS formalisms. A specification language permits the creation of coupled models, the initial configuration for the atomic models, and the creation of external events to be used during the simulation. The original version of CD++ permits the creation of bidimensional cellular automata, where the state of a cell has a binary or three-state value.The goal of this work is to introduce a set of extensions done to the CD++ tool [DW98], and others being implemented at present, so as to allow the execution of the models in parallel.2. CD++CD++ was defined building a class hierarchy in C++, using the basic concepts defined in [Zei84, Zei90]. Two basic abstract classes were defined: Models and Processors . The first are used to represent the behavior of the atomic and coupled models, while the second implement the simulation mechanisms. The Atomic class implements thebehavior of the atomic models. The Coupled-Model class implements the mechanisms of the coupled models. For the case of a cellular model, a special atomic model is used to represent to each cell. To do so, AtomicCell and CoupledCell are defined as subclasses of Atomic and Coupled respectively. AtomicCell extends the behavior of the atomic models, to define the functionality of the cell space. On the other hand, the CoupledCell class permits the management of a group of atomic cells.The Simulator and Coordinator classes manage the atomic and coupled models respectively. The Root-Coordinator class manages the global aspects of the simulation. It is directly related with the coupled model that has the higher level within the hierarchy. It is in charge to maintain the global time, and to start and stop the simulation process. In addition, it gets the output results.The simulation is based on the interchange of messages between the different processors. Each message contains information to identify the sender/receiver, the time of the event, and the content that consists in a port and a value for it. Different messages are used: X (which represents an external event), Y (represents the model output), * (represents an internal event), and done (indicating that the model has finished its task).The CD++ tool includes a specification language that allows describing the behavior of each cell of a cellular model. In addition, it allows to define the size of the cell space and their connection with other DEVS models (if they exist), the type of delay, the neighborhood, the border and the initial state of each cell. To do so, the theoretical definitions of the Cell-DEVS formalism are used.The specification of the behavior for a cell is defined using a set of rules. Each rule indicates the value for the cell's state if a condition is satisfied. The output of the model should be delayed by using a specified time. The BNF for the specification language is shown in the Appendix.The specification of the behavior for a cell is defined using a set of rules. Each rule indicates the value for the cell's state if a condition is satisfied. The output of the model should be delayed by using a specified time. If the condition is not valid, then the following rule will be evaluated (according to the order in that they were defined), repeating this process until a rule is satisfied, or until there are no more rules.In the latter case, an error will be raised, indicating this situation to the modeller, and aborting the simulation process. The occurrence of this error indicates that the model has been specified in incomplete form. The tool could also detect the existence of two or more rules with same condition but with different state value or delay, avoiding the creation of ambiguous models. In this situation, the simulation will be aborted. Also, when two different rules are evaluated satisfactorily and their result is the same, the modeller will be warned. 3. AN APPLICATION EXAMPLE: A HEATDIFFUSION MODELThis example consists of a surface represented by cellular automaton, where each cell contains a temperature. In each stage of the simulation, the temperature of the cell is calculated as the average of the values of the neighborhood. In addition, heat generator is connected to the cells (2, 2) and (5, 5), and permit the creation of temperatures in the range [24, 40] with uniform distribution. On the other hand, a generator of cold allow to create values in the range [10, 15] also with uniform distribution, and is connected to the cells (2, 8) and (8, 8). Both generators create values after x seconds, wherex follows an exponential distribution with mean 50 seconds.Figure 1.Coupling scheme of the heat diffusionmodel01 [top]02 components : surface generatorHeat@Generator generatorCold@Generator03 link : out@generatorHeat inputHeat@surface04 link : out@generatorCold inputCold@surface0506 [surface]07 type : cell08 width : 1009 height : 1010 delay : transport11 defaultDelayTime : 10012 border : wrapped13 neighbors : surface(-1,-1) surface(-1,0) surface(-1,1)14 neighbors : surface(0,-1) surface(0,0) surface(0,1)15 neighbors : surface(1,-1) surface(1,0) surface(1,1)16 initialvalue : 2417 in : inputHeat inputCold18 link : inputHeat in@surface(5,5)19 link : inputHeat in@surface(2,2)20 link : inputCold in@surface(8,8)21 link : inputCold in@surface(2,8)22 localtransition : heat-rule23 portInTransition : in@surface(5,5) setHeat24 portInTransition : in@surface(2,2) setHeat25 portInTransition : in@surface(8,8) setCold26 portInTransition : in@surface(2,8) setCold2728 [heat-rule]29 rule : { ((0,0) + (-1,-1) + (-1,0) + (-1,1) + (0,-1) + (0,1) + (1,-1) + (1,0) + (1,1)) / 9 } 10000 { t }3031 [setHeat]32 rule : { uniform(24,40) } 1000 { t }3334 [setCold]35 rule : { uniform(-10,15) } 1000 { t }3637 [generatorHeat]38 distribution : exponential39 mean : 5040 initial : 141 increment : 04243 [generatorCold]44 distribution : exponential45 mean : 5046 initial : 147 increment : 0The definition of the model using the language provided by the tool is showed in the figure 2. Thetop model and its components are defined between lines 1 and 4. Between lines 6 and 26, the model representing the surface is defined. It is composedby a cellular automata of 10x10 cells, having an initial temperature of 24° C. In the lines 28 and 29the local computation function is defined.Lines 31 and 32 define the function creating a temperature in the range [24, 40] with uniform distribution. Similarly, lines 34 and 35 define the function to create temperatures in the range [10, 15] with uniform distribution. Finally, the generators of cold and heat are defined between the lines 37 and 47. Here, the values are generate each x seconds, where x follows an exponential distribution with a mean of 50 seconds.The outputs generated by the simulation are givenin appendix. In the time 00:00:01:000 the generators of cold and heat produce changes in the input cells. At the time 00:00:05:041, the generatorof cold will produce a change in the state of the cells (2, 8) and (8, 8), establishing them the value 2.5 and -2.6 respectively.4. ABSTRACT SIMULATION MECHANISMIN ASYNCHRONOUS PDESTo improve the execution times in Cell-DEVS model execution, it has been proposed that the coordinators must be implemented to execute in parallel. In this case, multiple processes will execute the simulation simultaneously. Here, each processor of an available set will have an associatedflat coordinator. Asynchronous parallel discrete event simulation mechanisms will be used. The logical execution of processes associated with each processor will be synchronized by using optimist or pessimist approaches. In this way, the environment can be tailored to the application, achieving the best performance results for each case.Each coordinator will be coded as a logical process including three different event lists: one for local events, and the others for input and output links. The main synchronization mechanisms will be executed by three basic methods: Receive_message(), Transmit_Message() and Execute_Message(). The behavior of these methodswill differ depending on the chosen approach.Each coordinator will be in charge to choose the imminent cells to simulate, using the flatten procedure analyzed previously. The q-messages arriving to a coordinator can have a local or remote source. When these messages are processed, new,Y and done messages are created (using the translation mechanism explained previously). In this case, the coordinators must be changed to manage the coupling with the parallel simulation environment.Figure 3. Structure for the logical process definedfor cell spaces.If a new output event belongs to the local processor, it will be added directly to the next eventlist as it was shown in the previous procedures. Instead, if the destination cell does not belong tothe local processor, the message will be added intothe output queue and transmitted to other processor, where it will be received as an input message.The simulation can be splitted accordingly with the power of each processor, allowing the balance ofthe simulation load. The Map_Tasks() method willbe in charge of this procedure, by executing standard load distribution algorithms. This methodwill be in charge of saving a submap of cells into each coordinator, indicating to which processor does each part of the cell space belong. This information is recorded into the Processor_Mapping instance variable, and will be used to build the input and output links. The following sections will be devoted to analyze the behavior of the simulators for each of the proposed methods.4.1.1. Pessimist coordinators.This section is devoted to study the behavior of each of the methods associated with the conservative coordinators for a Cell-DEVS environment. The coordination mechanism should execute the following methods:Receive_Message() {Read a message arrived through an input link;Save it into the Event List;if the message correspond to the link withthe lower timestamp thenUnlock();}Execute_Message() {if the link with lower timestamp does nothave a message thenif Null Messages are used thenLookahead = LVT + d;Send a null message to the neighbors;endifLock(); /* The LP should be blockedwaiting the message with lower timestamp */ endifLVT = time of the first message in the event list; For each message in the queue withtimestamp = LVT docase kind of message do*: ReactionTo*Message();Send a done message to the parentcoordinator. If it belongs to otherprocessor, put it into thecorresponding output link.Q: ReactionToQMessage();Send a Y message to the parentcoordinator. If it belongs to otherprocessor, queue it into thecorresponding output link;Translate(Y, Q);/* Using the Neighborhood relationship */ Insert Q into the local events list;endcaseendfor}Transmit_Message() {For each element into an output port doSend the message through the corresponding link; }Figure 4. Pessimist coordination functions for Cell-DEVS parallel simulators.In this case, the method Receive_Message() must save a message into the input queue, recording its type, timestamp, origin/destination and value, according with the message type. When a message arrives, they are queued into the event list. The pair of Lock()/Unlock() methods are used to stop or reactivate the activity of the logical process. In this case, if a new message arrives with the lower timestamp, the logical process should be restarted. The method Execute_Message() must process all the events in the queue having timestamps lower than those of the local virtual time (LVT). This should be done only if there are messages into the input lists with timestamps higher than those of theLVT. If this is not the case, the logical process mustbe blocked. If the logical process is active, it muststart a cycle activating the methods ReactionToXMessage(), and ReactionTo*Message(), that will start the routinesof the flat coordinator.According with the strategy chosen to managedeadlocks, the method Recovery() (that detects andrecovers deadlocks) must be activated, or nullmessages must be sent. In this case, the lookaheadcomputation is easy, due that as each cell has afixed delay, the lookahead is equal to the delay. Themethod Transmit_Message() is activated when thecoordinator, analyzing its mapping (by using themethod Translate() that activates the Z function)decides that the message belongs to otherprocessor. To do so, the instance variableProcessor_Mapping is used.4.1.2. Optimist coordinatorsFor these processors, the methodsReceive_Message() and Transmit_Message() havethe same functionality that of the pessimist ones,and the difference can be found in theExecute_Message() method.Receive_Message() {Read a message arrived through an input link;Save it into the Event List;}Execute_Message() {if the first message in the event list has atimestamp lower than the LVT thenRollback();endifLVT = time of the first message;if LVT > GVT then /* Fossil collection */Delete all the events with timestamps lower tothe GVT to the Next-Events and Space lists;endifLVT = time of the first message in the event list;For each message in the queue whosetimestamp = LVT docase type of message do*: ReactionTo*Message();Send a done message to the parentcoordinator. If it belongs to otherprocessor, queue it into thecorresponding output link;Q: ReactionToQMessage();Send a Y message to the parent coordinator.If it belongs to other processor, queue itinto the corresponding output link;Translate(Y, Q);Insert Q into the local events queue;endcaseif a message identical to the present exists, andLAZY_ CANCELLATION is used thenif the message is the same except forthe value thenGenerate an antimessage using the previous;Queue the present;endifendifendforUpdate the Next-Events and Spaces lists with thenew status;}Rollback() {LVT = time for the straggler;Search into the Next-events and Space liststhe first element whose value is smallerthan LVT;Cells = Space.Cells;Next_Events = Next-Events.list;case CANCELLATION doLAZY: nothing;AGGRESIVE: generate an antimessageinto the output queue;endcaseDelete the elements whose timestampis smaller than the LVT from the outputqueue;}Figure 5. Optimistic coordination functions forCell-DEVS models.In this case, the first messages of the events list are used, analyzing if the arrived event is a straggler or not. If this is the case, the event list should be restored to the instant for the straggler, as the cell space is. To do so, a list keeping all the Next-Events and other with all the cell spaces are kept. The straggler generates antimessages, and restores the previous state.If aggressive cancellation is used, the Transmit_Message() method should be activated when the coordinator, analyzing the mapping, decides that the message belongs to a coordinator ina different processor. If lazy cancellation is used, the method deletes safe elements from the output queue, and antimessages are not transmitted for messages with the same contents. 5. CONCLUSIONThis work introduced an extension to the tool CD++ used for the modeling and simulation of Cell-DEVS models. This formalism allows hierarchical construction of the models, which improves the development, checking and maintenance phases. The extensions introduced to the tool allow to represent new models in other domains for the sate variables. It also offers the possibility to use probabilistic functions, which permits the creation of stochastic models.The parallel execution of these models is being considered at present. A simulation mechanism was presented for this kind of models. In addition, a new extension to the flat coordination mechanism was introduced. The formalism entitles the definition of complex cell-shaped models using a high level specification language. In this way, the construction of the simulators is improved, enhancing their safety and development costs. Besides, the parallel execution allows performance improvements without adding extra costs in development or maintenance.REFERENCES[BBW98] BARYLKO, A.; BEYOGLONIAN, J.; WAINER, G. "CD++: a tool to develop binary Cell-DEVS models" (in Spanish). Proceedings of the XXII Latin-American Conference on Informatics. Quito, Ecuador. 1998.[Gar70] GARDNER, M. “The Fantastic Combinations of John Conway’s New Solitaire Game ‘Life’ ”. Scientific American, 23 (4), 1970, pp. 120-123.[GM76] N.Giambiasi, A.Miara "SILOG: A practical tool for digital logic circuit simulation" 16th D.A.C San Diego , 1976[GG96] GHOSH, S.; GIAMBIASI, N. "On the need for consistency between the VHDL language constructions and the underliying hardware design". SCS ESM '96. pp. 562-567.[WG98] WAINER, G.; GIAMBIASI, N. "Specification, modeling and simulation of timed Cell-DEVS spaces". Technical Report n.: 98-007. Departamento de Computación. Facultad de Ciencias Exactas y Naturales. Universidad de Buenos Aires. Submitted for publication. 1998. [RW99] RODRIGUEZ, D.; WAINER, G. "Extensions to the CD++ tool". Submitted for publication. 1999.[Zei76] ZEIGLER, B. “Theory of Modelling and Simulation”. Wiley, N.Y. 1976.[Zei84] ZEIGLER, B. "Multifaceted Modeling and discrete event simulation". Academic Press, 1984. [Zei90] ZEIGLER, B. "Object-oriented simulation with hierarchical modular models". Academic Press, 1990.APPENDIX. Simulation resultsTime: 00:00:00:0000 1 2 3 4 5 6 7 8 9 +--------------------------------------------------+ 0| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 1| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 2| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 3| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 4| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 5| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 6| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 7| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 8| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 9| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| +--------------------------------------------------+Time: 00:00:02:0000 1 2 3 4 5 6 7 8 9 +--------------------------------------------------+ 0| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 1| 24.0 25.3 25.3 25.3 24.0 24.0 24.0 20.6 20.6 20.6| 2| 24.0 25.3 25.3 25.3 24.0 24.0 24.0 20.6 20.6 20.6| 3| 24.0 25.3 25.3 25.3 24.0 24.0 24.0 20.6 20.6 20.6| 4| 24.0 24.0 24.0 24.0 25.7 25.7 25.7 24.0 24.0 24.0| 5| 24.0 24.0 24.0 24.0 25.7 25.7 25.7 24.0 24.0 24.0| 6| 24.0 24.0 24.0 24.0 25.7 25.7 25.7 24.0 24.0 24.0| 7| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 20.9 20.9 20.9| 8| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 20.9 20.9 20.9| 9| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 20.9 20.9 20.9| +--------------------------------------------------+. . .Time: 00:00:01:0000 1 2 3 4 5 6 7 8 9 +--------------------------------------------------+ 0| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 1| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 2| 24.0 24.0 35.8 24.0 24.0 24.0 24.0 24.0 -6.3 24.0| 3| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| 4| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0|5| 24.0 24.0 24.0 24.0 24.0 39.5 24.0 24.0 24.0 24.0|6| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0|7| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0|8| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 -4.0 24.0| 9| 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0 24.0| +--------------------------------------------------+. . .Time: 00:00:05:0410 1 2 3 4 5 6 7 8 9 +--------------------------------------------------+ 0| 23.3 23.9 24.3 24.3 24.1 23.7 23.1 22.6 22.3 22.6| 1| 23.4 24.1 24.5 24.5 24.2 23.8 23.2 22.6 22.3 22.7| 2| 23.4 24.2 24.6 24.6 24.4 23.9 23.3 22.6 2.5 22.7| 3| 23.5 24.2 24.6 24.7 24.6 24.3 23.7 23.1 22.7 23.0| 4| 23.7 24.1 24.4 24.7 24.7 24.6 24.1 23.6 23.2 23.3| 5| 23.7 24.0 24.3 24.6 24.8 24.7 24.4 23.9 23.5 23.5| 6| 23.5 23.9 24.1 24.4 24.6 24.5 24.1 23.6 23.3 23.3| 7| 23.3 23.8 24.0 24.2 24.3 24.2 23.7 23.1 22.8 22.9| 8| 23.2 23.7 24.0 24.1 24.1 23.8 23.3 22.7 -2.6 22.6| 9| 23.2 23.8 24.1 24.1 24.0 23.7 23.2 22.6 22.4 22.6| +--------------------------------------------------+Simulation results of the heat diffusion modelAPPENDIX. CD++ SPECIFICATION LANGUAGE.RULELIST = RULE | RULELIST RULE = RESULT RESULT { BOOLEXP } RESULT = CONSTANT | { REALEXP }BOOLEXP = BOOL | (BOOLEXP ) | REALRELEXP | not BOOLEXP | BOOLEXP OP_BOOL BOOLEXPOP_BOOL =and | or | xor | imp | eqv REALRELEXP = REALEXP OP_REL REALEXP | COND_REAL_FUNC (REALEXP ) REALEXP = IDREF | (REALEXP ) | REALEXP OPER REALEXP IDREF = CELLREF | CONSTANT | FUNCTION | portValue(PORTNAME ) CONSTANT = INT | REAL | CONSTFUNC | ? FUNCTION = UNARY_FUNC (REALEXP ) | WITHOUT_PARAM_FUNC | BINARY_FUNC (REALEXP , REALEXP ) | if(BOOLEXP , REALEXP , REALEXP ) | ifu(BOOLEXP , REALEXP , REALEXP , REALEXP )CELLREF =(INT , INT ) BOOL = t | f | ? OP_REL = != | = | > | < | >= | <= OPER =+ | - | * | / INT = [SIGN] DIGIT {DIGIT} REAL = INT | [SIGN] {DIGIT}.DIGIT {DIGIT}SIGN =+ | - DIGIT =0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 PORTNAME =thisPort | STRING STRING = LETTER {LETTER}LETTER =a |b |c |...| z | A | B | C |...| Z CONSTFUNC = pi | e | inf | grav | accel | light | planck | avogadro | faraday | rydberg | euler_gamma | bohr_radius | boltzmann | bohr_magneton | golden | catalan | amu | electron_charge | pem | ideal_gas | stefan_boltzmann | proton_mass | electron_mass | neutron_massWITHOUT_PARAM_FUNC = truecount | falsecount | u ndefcount | time | random | randomSign UNARY_FUNC = abs | acos | acosh | asin | asinh | atan | atanh | cos | sec | sech | exp | cosh | fact | fractional | ln | log | round | cotan | cosec | cosech | sign | sin | sinh | statecount | sqrt | tan | tanh | trunc | truncUpper | poisson | exponential | randInt | chi | asec | acotan | asech | acosech | nextPrime | radToDeg | degToRad | nth_prime | acotanh | CtoF | CtoK | KtoC | KtoF | FtoC | FtoKBINARY_FUNC = comb | logn | max | min | power | remainder | root | beta | gamma | lcm | gcd | normal | f | uniform | binomial | rectToPolar_r | rectToPolar_angle | polarToRect_x | hip | polarToRect_yCOND_REAL_FUNC = even | odd | isInt | isPrime | isUndefined。

Radio-immunotherapy

Commentchemotherapy is insuﬃ cient to change clinical practice without more clinically meaningful improvements, or without improving long-term cure rates or delaying metastatic or symptom progression. At present, docetaxel should be reserved for metastatic disease either in the castration-sensitive setting, such as was assessed in CHAARTE D and STAMPE DE , or for castration-resistant disease. Ongoing trials, such as the RTOG 0521 trial of external beam radiotherapy and long-term androgen-deprivation therapy with or without adjuvant docetaxel (NCT00288080), and the CALGB/ALLIANCE 90203 (PUNCH) trial (NCT00430183) of neoadjuvant docetaxel and androgen-deprivation therapy before radical prostatectomy versus immediate surgery should further clarify the optimal timing of docetaxel. However, GE TUG-12 adds to the growing evidence challenging the dogma that chemotherapy is ineﬀ ective in prostate cancer, and suggests that early docetaxel can safely delay progression. Whether this delay in can increase life expectancy will need to be addressed in larger well-powered studies.R adio-immunotherapy: the focused beam expandsLocal treatment of tumour masses by ionising radiation with additional immune stimulation could result in systemic and immune-mediated anti-tumour responses. The abscopal eﬀ ect of radiotherapy for malignant lymphoma was described by Nobler 1 in 1969 and has had a renaissance in the last decade when immunotherapies made their way into the clinics.Preclinical work has shown that, especially in melanoma, the combination of radiotherapy with immune checkpoint inhibitors such as anti-CTLA-4 and anti-PDL-1 antibodies results in CD8+ T-cell mediated anti-tumour responses far outside of the irradiated area.2 And in a case report of a patient with melanoma treated with radiotherapy and ipilumumab, immune alterations—eg, increased numbers of antigen-presenting cells and decreased numbers of myeloid-derived suppressor cells—were noted in the peripheral blood.3 Combining local radiotherapy with further immune stimulation eﬀ ectively turns the tumour into its own vaccine.4In The Lancet Oncology , Encouse Golden and colleagues 5have explored the interaction between radiotherapy and immunotherapy for the ﬁ rst time in other solid metastatic tumours such as non-small-cell lung cancer and breast cancer. Combination of radiotherapy with immunotherapy stimulating the innate and adaptive immune system has already been shown to be feasible, but abscopal responses have not been assessed. Local tumour responses can be enhanced by combining radiochemotherapy with immunotherapy. It was believed for a long time that no synergy exists because of the immune suppressive properties of the classical tumour therapies. However, recent data clearly show that radiotherapy and chemotherapy can be combined with immunotherapy.6In Golden and colleagues’ study,5 the combination of granulocyte-macrophage colony-stimulating factor (GM-CSF) with radiochemotherapy resulted in abscopal responses in four (22%) of 18 patients with non-small-cell lung cancer and ﬁ ve (36%) of 14 patients with breast cancer. These ﬁ ndings emphasise that systemicanti-tumour immunity can be induced by renderingthe tumour cells immunogenic. Radiotherapy alone is capable of enhancing MHC-I surface expression on tumour cells, releasing danger signals, and also broadensAndrew J ArmstrongDUMC Box 103861, Duke Cancer Institute, Divisions of Medical Oncology and Urology, Duke University, Durham, NC 27710, USA andrew.armstrong@I have received research grant support to my institution from Sanoﬁ Aventis.1Petrylak DP, Tangen CM, Hussain MH, et al. Docetaxel and estramustinecompared with mitoxantrone and prednisone for advanced refractory prostate cancer. N Engl J Med 2004; 351: 1513–20.2Tannock IF, Osoba D, Stockler MR, et al. Chemotherapy with mitoxantrone plus prednisone or prednisone alone for symptomatic hormone-resistant prostate cancer: a Canadian randomized trial with palliative end points. J Clin Oncol 1996; 14: 1756–64.3Sweeney C, Chen Y-H, Carducci MA, et al. Impact on overall survival (OS) with chemohormonal therapy versus hormonal therapy for hormone-sensitive newly metastatic prostate cancer (mPrCa): an ECOG-led phase III randomized trial. Proc Am Soc Clin Oncol 2014; 32 (suppl): abtr LBA2. 4Fizazi K, Faivre L, Lesaunier F, et al. Androgen deprivation therapy plus docetaxel and estramustine versus androgen deprivation therapy alone for high-risk localised prostate cancer (GETUG 12): a phase 3 randomised controlled trial. Lancet Oncol 2015; published online May 29. /10.1016/S1470-2045(15)00011-X.5Daskivich TJ, Fan KH, Koyama T, et al. Prediction of long-term other-cause mortality in men with early-stage prostate cancer: results from the Prostate Cancer Outcomes Study. Urology 2015; 85: 92–100.6Kelly WK, Halabi S, Carducci M, et al. Randomized, double-blind,placebo-controlled phase III trial comparing docetaxel and prednisone with or without bevacizumab in men with metastatic castration-resistant prostate cancer: CALGB 90401. J Clin Oncol 2012; 30: 1534–40.7Michaelson MD, Oudard S, Ou YC, et al. Randomized, placebo-controlled, phase III trial of sunitinib plus prednisone versus prednisone alone in progressive, metastatic, castration-resistant prostate cancer. J Clin Oncol 2014; 32: 76–82.Published OnlineJune 19, 2015/10.1016/S1470-2045(15)00055-8See Articles page 795L a r r y M u l v e h i l l /S c i e n c e P h o t o L i b r a r yCommentthe peptide pool.7 However, it is obviously not suﬃ cient to induce abscopal responses, as proven by clinical observation in patients receiving radiochemotherapy. The immunological balance has to be pushed towards anti-tumour immunity, and addition of immunotherapy is the solution.8 Flt3-L and GM-CSF are both strong stimulators of dendritic cells.The Sipuleucel-T vaccine study 9 for prostate cancer shows commonalities with the study of Golden and colleagues.5Since no control group consisting of only GM-CSF was included in the study of Kantoﬀ and colleagues 9the observed increase of median survivalwith the vaccine might also be due to GM-CSF and not the antigenic peptide.The abscopal anti-tumour responses were accompanied by stable low concentrations of neutrophils. A challenge for future research is to determine whether these stable concentrations, together with increasing amounts of dendritic cells and CD8+ T cells, might be predictors forradio-immunotherapy-induced abscopal responses. Further, correlations with quality of life of respondingpatients should be assessed, since metastatic lesions are the speciﬁ c targets of such immune responses. Whetherbone and spinal metastases respond diﬀ erently comparedwith visceral metastases also needs to be explored, since their accessibility by the immune system is reduced.In patients with a high metastatic tumour burden, the induction of abscopal anti-tumour reactions could be awelcome means to palliate the disease when the immune evasion of the tumour is at a maximum. The continuing rapid development of more precise and sophisticated accelerators for delivery of ionising irradiation should allow us to investigate the immunological consequences of classical fractionated versus hypofractionated or stereotactic high-dose delivery. Current preclinical dataare not conclusive. For instance, the combination of anti-CTLA-4 immunotherapy with fractionated butnot hypofractionated radiotherapy induced abscopalanti-tumour responses in a mouse breast cancer model,10 whereas in a mouse model of melanoma, ablative radiotherapy was particularly immunogenic.11The study by Golden and colleagues shows, for the ﬁ rst time, that abscopal anti-tumour responses can be induced in solid metastatic tumours by combining classical radiochemotherapy with simple immunotherapy. Nevertheless, randomised trials are needed for detailed assessment of the eﬃ cacy of radio-immunotherapy ininducing abscopal responses, in particular assessing the best combinations and chronology of treatments.Benjamin Frey, Udo S GaiplDepartment of Radiation Oncology, University Hospital Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg,Universitaetsstrasse 27, 91054 Erlangen, Germanybenjamin.frey@uk-erlangen.de We declare no competing interests.1 Nobler MP. The abscopal eﬀ ect in malignant lymphoma and its relationshipto lymphocyte circulation. Radiology 1969; 93: 410–12.2 Twyman-Saint Victor C, Rech AJ, Maity A, et al. Radiation and dualcheckpoint blockade activate non-redundant immune mechanisms incancer. Nature 2015; 520: 373–77.3 Postow MA, Callahan MK, Barker CA, et al. Immunologic correlates of theabscopal eﬀ ect in a patient with melanoma. N Engl J Med 2012; 366: 925–31.4 Frey B, Rubner Y, Kulzer L, et al. Antitumor immune responses induced by ionizing irradiation and further immune stimulation. Cancer Immunol Immunother 2014; 63: 29–36.5 Golden EB, Chhabra A, Chachoua A, et al. Local radiotherapy and granulocyte-macrophage colony-stimulating factor to generate abscopal responses in patients with metastatic solid tumours: a proof-of-principle trial. Lancet Oncol 2015; published June 19. /10.1016/S1470-2045(15)00054-6.6 van der Sluis TC, van Duikeren S, Huppelschoten S, et al. Vaccine-inducedtumor necrosis factor-producing T cells synergize with cisplatin to promote tumor cell death. Clin Cancer Res 2015; 21: 781–94.7 Hodge JW, Ardiani A, Farsaci B, Kwilas AR, Gameiro SR. The tipping point forcombination therapy: cancer vaccines with radiation, chemotherapy, or targeted small molecule inhibitors. Semin Oncol 2012; 39: 323–39.8 Demaria S, Ng B, Devitt ML, et al. Ionizing radiation inhibition of distantuntreated tumors (abscopal eﬀ ect) is immune mediated.Int J Radiat Oncol Biol Phys 2004; 58: 862–70.9 Kantoﬀ PW, Higano CS, Shore ND, et al. Sipuleucel-T immunotherapy forcastration-resistant prostate cancer. N Engl J Med 2010; 363: 411–22.10 Dewan MZ, Galloway AE, Kawashima N, et al. Fractionated but not single-dose radiotherapy induces an immune-mediated abscopal eﬀ ect when combined with anti-CTLA-4 antibody. Clin Cancer Res 2009; 15: 5379–88.11 Lee Y, Auh SL, Wang Y, et al. Therapeutic eﬀ ects of ablative radiation onlocal tumor require CD8+ T cells: changing strategies for cancer treatment. Blood 2009; 114: 589–95.Hepatocellular carcinoma surveillance with miRNAsHepatocellular carcinoma is one of the most lethal cancers worldwide.1 Chronic liver disease is the main risk factor for its development and hepatocellular carcinoma is the most frequent cause of death in this population.2 The likelihood of long-term disease-free survivalincreases when the cancer is diagnosed at an early, asymptomatic stage, when potential curative treatments are feasible.1arly diagnosis before appearance of symptoms is only feasible in a surveillance programme and several observational and cost-eﬀ ectiveness studiesPublished Online June 16, 2015/10.1016/S1470-2045(15)00014-5See Articles page 804。

室内pm2.5浓度标准及其控制策略

室内PM2.5浓度标准及其控制策略朱梅文远高（武汉科技大学环境工程）摘要：室内颗粒物中占有很大比例的PM2.5由于比表面积大，对多种有机物具有较强的吸附能力，可以直接进入肺泡，增加年总死亡率、心肺疾病死亡率以及肺癌的死亡率，进而对人体产生全方位的影响。

结合各国/组织的PM2.5标准及室内PM2.5检测标准可以为《室内空气质量标准》中PM2.5标准值的预测给予科学依据。

通过研究PM2.5的形成机理，分析影响室内PM2.5浓度的因素,针对采用不同空调系统的房间可以提出其室内PM2.5浓度的控制策略，从而达到进一步改善室内环境质量的目的，让人们拥有一个更加舒适健康的室内环境。

关键词：PM2.5形成机理浓度标准控制策略。

The concentration standard and control techniques of indoor PM2.5Zhu Mei Wen YuangaoAbstract: The PM2.5, which occupies a large proportion in indoor particulate matters, due to its large specific surface area and strong adsorption capacity for a variety of organic compounds , can directly access the alveoli and increase the total mortality, heart and lung disease mortality and the mortality of lung cancer. So it can make a full range of impact on the human body. The nations / organizations indoor PM2.5 standards and the testing standard can provide scientific basis for predicting PM2.5 standard values in "Indoor Air Quality Standards". Through the research on the formation mechanism of PM2.5 and analyzing the factors affecting indoor PM2.5 concentrations, we can put forward control strategies on indoor PM2.5 for rooms using different air-conditioning systems and further improve the indoor environmental quality. For all of these, people can have a more comfortable and healthier environment.Keywords:PM2.5Formation Mechanism Concentration standard Control techniques1 前言2013年1月9日以来，全国中东部地区大范围的雾霾天气让颗粒污染物PM2.5从一个专业术语变成了切实的感受。

Understanding individual human mobility patterns

a r X i v :0806.1256v 1 [p h y s i c s .s o c -p h ] 7 J u n 2008Understanding individual human mobility patternsMarta C.Gonz´a lez,1,2C´e sar A.Hidalgo,1and Albert-L´a szl´o Barab´a si 1,2,31Center for Complex Network Research and Department of Physics and Computer Science,University of Notre Dame,Notre Dame IN 46556.2Center for Complex Network Research and Department of Physics,Biology and Computer Science,Northeastern University,Boston MA 02115.3Center for Cancer Systems Biology,Dana Farber Cancer Institute,Boston,MA 02115.(Dated:June 7,2008)Despite their importance for urban planning [1],trafﬁc forecasting [2],and the spread of biological [3,4,5]and mobile viruses [6],our understanding of the basic laws govern-ing human motion remains limited thanks to the lack of tools to monitor the time resolved location of individuals.Here we study the trajectory of 100,000anonymized mobile phone users whose position is tracked for a six month period.We ﬁnd that in contrast with the random trajectories predicted by the prevailing L´e vy ﬂight and random walk models [7],human trajectories show a high degree of temporal and spatial regularity,each individual being characterized by a time independent characteristic length scale and a signiﬁcant prob-ability to return to a few highly frequented locations.After correcting for differences in travel distances and the inherent anisotropy of each trajectory,the individual travel patterns collapse into a single spatial probability distribution,indicating that despite the diversity of their travel history,humans follow simple reproducible patterns.This inherent similarity in travel patterns could impact all phenomena driven by human mobility,from epidemic prevention to emergency response,urban planning and agent based modeling.Given the many unknown factors that inﬂuence a population’s mobility patterns,ranging from means of transportation to job and family imposed restrictions and priorities,human trajectories are often approximated with various random walk or diffusion models [7,8].Indeed,early mea-surements on albatrosses,bumblebees,deer and monkeys [9,10]and more recent ones on marine predators [11]suggested that animal trajectory is approximated by a L´e vy ﬂight [12,13],a random walk whose step size ∆r follows a power-law distribution P (∆r )∼∆r −(1+β)with β<2.While the L´e vy statistics for some animals require further study [14],Brockmann et al.[7]generalized this ﬁnding to humans,documenting that the distribution of distances between consecutive sight-ings of nearly half-million bank notes is fat tailed.Given that money is carried by individuals, bank note dispersal is a proxy for human movement,suggesting that human trajectories are best modeled as a continuous time random walk with fat tailed displacements and waiting time dis-tributions[7].A particle following a L´e vyﬂight has a signiﬁcant probability to travel very long distances in a single step[12,13],which appears to be consistent with human travel patterns:most of the time we travel only over short distances,between home and work,while occasionally we take longer trips.Each consecutive sightings of a bank note reﬂects the composite motion of two or more indi-viduals,who owned the bill between two reported sightings.Thus it is not clear if the observed distribution reﬂects the motion of individual users,or some hitero unknown convolution between population based heterogeneities and individual human trajectories.Contrary to bank notes,mo-bile phones are carried by the same individual during his/her daily routine,offering the best proxy to capture individual human trajectories[15,16,17,18,19].We used two data sets to explore the mobility pattern of individuals.Theﬁrst(D1)consists of the mobility patterns recorded over a six month period for100,000individuals selected randomly from a sample of over6million anonymized mobile phone users.Each time a user initiates or receives a call or SMS,the location of the tower routing the communication is recorded,allowing us to reconstruct the user’s time resolved trajectory(Figs.1a and b).The time between consecutive calls follows a bursty pattern[20](see Fig.S1in the SM),indicating that while most consecutive calls are placed soon after a previous call,occasionally there are long periods without any call activity.To make sure that the obtained results are not affected by the irregular call pattern,we also study a data set(D2)that captures the location of206mobile phone users,recorded every two hours for an entire week.In both datasets the spatial resolution is determined by the local density of the more than104mobile towers,registering movement only when the user moves between areas serviced by different towers.The average service area of each tower is approximately3km2 and over30%of the towers cover an area of1km2or less.To explore the statistical properties of the population’s mobility patterns we measured the dis-tance between user’s positions at consecutive calls,capturing16,264,308displacements for the D1and10,407displacements for the D2datasets.Weﬁnd that the distribution of displacements over all users is well approximated by a truncated power-lawP(∆r)=(∆r+∆r0)−βexp(−∆r/κ),(1)withβ=1.75±0.15,∆r0=1.5km and cutoff valuesκ|D1=400km,andκ|D2=80km(Fig.1c,see the SM for statistical validation).Note that the observed scaling exponent is not far fromβB=1.59observed in Ref.[7]for bank note dispersal,suggesting that the two distributions may capture the same fundamental mechanism driving human mobility patterns.Equation(1)suggests that human motion follows a truncated L´e vyﬂight[7].Yet,the observed shape of P(∆r)could be explained by three distinct hypotheses:A.Each individual follows a L´e vy trajectory with jump size distribution given by(1).B.The observed distribution captures a population based heterogeneity,corresponding to the inherent differences between individuals.C.A population based heterogeneity coexists with individual L´e vy trajectories,hence(1)represents a convolution of hypothesis A and B.To distinguish between hypotheses A,B and C we calculated the radius of gyration for each user(see Methods),interpreted as the typical distance traveled by user a when observed up to time t(Fig.1b).Next,we determined the radius of gyration distribution P(r g)by calculating r g for all users in samples D1and D2,ﬁnding that they also can be approximated with a truncated power-lawP(r g)=(r g+r0g)−βr exp(−r g/κ),(2) with r0g=5.8km,βr=1.65±0.15andκ=350km(Fig.1d,see SM for statistical validation). L´e vyﬂights are characterized by a high degree of intrinsic heterogeneity,raising the possibility that(2)could emerge from an ensemble of identical agents,each following a L´e vy trajectory. Therefore,we determined P(r g)for an ensemble of agents following a Random Walk(RW), L´e vy-Flight(LF)or Truncated L´e vy-Flight(T LF)(Figure1d)[8,12,13].Weﬁnd that an en-semble of L´e vy agents display a signiﬁcant degree of heterogeneity in r g,yet is not sufﬁcient to explain the truncated power law distribution P(r g)exhibited by the mobile phone users.Taken together,Figs.1c and d suggest that the difference in the range of typical mobility patterns of indi-viduals(r g)has a strong impact on the truncated L´e vy behavior seen in(1),ruling out hypothesis A.If individual trajectories are described by a LF or T LF,then the radius of gyration should increase in time as r g(t)∼t3/(2+β)[21,22]while for a RW r g(t)∼t1/2.That is,the longer we observe a user,the higher the chances that she/he will travel to areas not visited before.To check the validity of these predictions we measured the time dependence of the radius of gyration for users whose gyration radius would be considered small(r g(T)≤3km),medium(20<r g(T)≤30km)or large(r g(T)>100km)at the end of our observation period(T=6months).Theresults indicate that the time dependence of the average radius of gyration of mobile phone users is better approximated by a logarithmic increase,not only a manifestly slower dependence than the one predicted by a power law,but one that may appear similar to a saturation process(Fig.2a and Fig.S4).In Fig.2b,we have chosen users with similar asymptotic r g(T)after T=6months,and measured the jump size distribution P(∆r|r g)for each group.As the inset of Fig.2b shows,users with small r g travel mostly over small distances,whereas those with large r g tend to display a combination of many small and a few larger jump sizes.Once we rescale the distributions with r g(Fig.2b),weﬁnd that the data collapses into a single curve,suggesting that a single jump size distribution characterizes all users,independent of their r g.This indicates that P(∆r|r g)∼r−αg F(∆r/r g),whereα≈1.2±0.1and F(x)is an r g independent function with asymptotic behavior F(x<1)∼x−αand rapidly decreasing for x≫1.Therefore the travel patterns of individual users may be approximated by a L´e vyﬂight up to a distance characterized by r g. Most important,however,is the fact that the individual trajectories are bounded beyond r g,thus large displacements which are the source of the distinct and anomalous nature of L´e vyﬂights, are statistically absent.To understand the relationship between the different exponents,we note that the measured probability distributions are related by P(∆r)= ∞0P(∆r|r g)P(r g)dr g,whichsuggests(see SM)that up to the leading order we haveβ=βr+α−1,consistent,within error bars, with the measured exponents.This indicates that the observed jump size distribution P(∆r)is in fact the convolution between the statistics of individual trajectories P(∆r g|r g)and the population heterogeneity P(r g),consistent with hypothesis C.To uncover the mechanism stabilizing r g we measured the return probability for each indi-vidual F pt(t)[22],deﬁned as the probability that a user returns to the position where it was ﬁrst observed after t hours(Fig.2c).For a two dimensional random walk F pt(t)should follow ∼1/(t ln(t)2)[22].In contrast,weﬁnd that the return probability is characterized by several peaks at24h,48h,and72h,capturing a strong tendency of humans to return to locations they visited before,describing the recurrence and temporal periodicity inherent to human mobility[23,24].To explore if individuals return to the same location over and over,we ranked each location based on the number of times an individual was recorded in its vicinity,such that a location with L=3represents the third most visited location for the selected individual.Weﬁnd that the probability ofﬁnding a user at a location with a given rank L is well approximated by P(L)∼1/L, independent of the number of locations visited by the user(Fig.2d).Therefore people devote mostof their time to a few locations,while spending their remaining time in5to50places,visited with diminished regularity.Therefore,the observed logarithmic saturation of r g(t)is rooted in the high degree of regularity in their daily travel patterns,captured by the high return probabilities(Fig.2b) to a few highly frequented locations(Fig.2d).An important quantity for modeling human mobility patterns is the probabilityΦa(x,y)toﬁnd an individual a in a given position(x,y).As it is evident from Fig.1b,individuals live and travel in different regions,yet each user can be assigned to a well deﬁned area,deﬁned by home and workplace,where she or he can be found most of the time.We can compare the trajectories of different users by diagonalizing each trajectory’s inertia tensor,providing the probability ofﬁnding a user in a given position(see Fig.3a)in the user’s intrinsic reference frame(see SM for the details).A striking feature ofΦ(x,y)is its prominent spatial anisotropy in this intrinsic reference frame(note the different scales in Fig3a),and weﬁnd that the larger an individual’s r g the more pronounced is this anisotropy.To quantify this effect we deﬁned the anisotropy ratio S≡σy/σx, whereσx andσy represent the standard deviation of the trajectory measured in the user’s intrinsic reference frame(see SM).Weﬁnd that S decreases monotonically with r g(Fig.3c),being well approximated with S∼r−ηg,forη≈0.12.Given the small value of the scaling exponent,other functional forms may offer an equally goodﬁt,thus mechanistic models are required to identify if this represents a true scaling law,or only a reasonable approximation to the data.To compare the trajectories of different users we remove the individual anisotropies,rescal-ing each user trajectory with its respectiveσx andσy.The rescaled˜Φ(x/σx,y/σy)distribution (Fig.3b)is similar for groups of users with considerably different r g,i.e.,after the anisotropy and the r g dependence is removed all individuals appear to follow the same universal˜Φ(˜x,˜y)prob-ability distribution.This is particularly evident in Fig.3d,where we show the cross section of ˜Φ(x/σ,0)for the three groups of users,ﬁnding that apart from the noise in the data the curves xare indistinguishable.Taken together,our results suggest that the L´e vy statistics observed in bank note measurements capture a convolution of the population heterogeneity(2)and the motion of individual users.Indi-viduals display signiﬁcant regularity,as they return to a few highly frequented locations,like home or work.This regularity does not apply to the bank notes:a bill always follows the trajectory of its current owner,i.e.dollar bills diffuse,but humans do not.The fact that individual trajectories are characterized by the same r g-independent two dimen-sional probability distribution˜Φ(x/σx,y/σy)suggests that key statistical characteristics of indi-vidual trajectories are largely indistinguishable after rescaling.Therefore,our results establish the basic ingredients of realistic agent based models,requiring us to place users in number propor-tional with the population density of a given region and assign each user an r g taken from the observed P(r g)ing the predicted anisotropic rescaling,combined with the density function˜Φ(x,y),whose shape is provided as Table1in the SM,we can obtain the likelihood ofﬁnding a user in any location.Given the known correlations between spatial proximity and social links,our results could help quantify the role of space in network development and evolu-tion[25,26,27,28,29]and improve our understanding of diffusion processes[8,30].We thank D.Brockmann,T.Geisel,J.Park,S.Redner,Z.Toroczkai and P.Wang for discus-sions and comments on the manuscript.This work was supported by the James S.McDonnell Foundation21st Century Initiative in Studying Complex Systems,the National Science Founda-tion within the DDDAS(CNS-0540348),ITR(DMR-0426737)and IIS-0513650programs,and the U.S.Ofﬁce of Naval Research Award N00014-07-C.Data analysis was performed on the Notre Dame Biocomplexity Cluster supported in part by NSF MRI Grant No.DBI-0420980.C.A.Hi-dalgo acknowledges support from the Kellogg Institute at Notre Dame.Supplementary Information is linked to the online version of the paper at /nature.Author Information Correspondence and requests for materials should be addressed to A.-L.B.(e-mail:alb@)[1]Horner,M.W.&O’Kelly,M.E.S Embedding economies of scale concepts for hub networks design.Journal of Transportation Geography9,255-265(2001).[2]Kitamura,R.,Chen,C.,Pendyala,R.M.&Narayaran,R.Micro-simulation of daily activity-travelpatterns for travel demand forecasting.Transportation27,25-51(2000).[3]Colizza,V.,Barrat,A.,Barth´e l´e my,M.,Valleron,A.-J.&Vespignani,A.Modeling the WorldwideSpread of Pandemic Inﬂuenza:Baseline Case and Containment Interventions.PLoS Medicine4,095-0110(2007).[4]Eubank,S.,Guclu,H.,Kumar,V.S.A.,Marathe,M.V.,Srinivasan,A.,Toroczkai,Z.&Wang,N.Controlling Epidemics in Realistic Urban Social Networks.Nature429,180(2004).[5]Hufnagel,L.,Brockmann,D.&Geisel,T.Forecast and control of epidemics in a globalized world.Proceedings of the National Academy of Sciences of the United States of America101,15124-15129 (2004).[6]Kleinberg,J.The wireless epidemic.Nature449,287-288(2007).[7] D.Brockmann,D.,Hufnagel,L.&Geisel,T.The scaling laws of human travel.Nature439,462-465(2006).[8]Havlin,S.&ben-Avraham,D.Diffusion in Disordered Media.Advances in Physics51,187-292(2002).[9]Viswanathan,G.M.,Afanasyev,V.,Buldyrev,S.V.,Murphy,E.J.,Prince,P.A.&Stanley,H.E.L´e vyFlight Search Patterns of Wandering Albatrosses.Nature381,413-415(1996).[10]Ramos-Fernandez,G.,Mateos,J.L.,Miramontes,O.,Cocho,G.,Larralde,H.&Ayala-Orozco,B.,L´e vy walk patterns in the foraging movements of spider monkeys(Ateles geoffroyi).Behavioral ecol-ogy and Sociobiology55,223-230(2004).[11]Sims D.W.et al.Scaling laws of marine predator search behaviour.Nature451,1098-1102(2008).[12]Klafter,J.,Shlesinger,M.F.&Zumofen,G.Beyond Brownian Motion.Physics Today49,33-39(1996).[13]Mantegna,R.N.&Stanley,H.E.Stochastic Process with Ultraslow Convergence to a Gaussian:TheTruncated L´e vy Flight.Physical Review Letters73,2946-2949(1994).[14]Edwards,A.M.,Phillips,R.A.,Watkins,N.W.,Freeman,M.P.,Murphy,E.J.,Afanasyev,V.,Buldyrev,S.V.,da Luz,M.G.E.,Raposo,E.P.,Stanley,H.E.&Viswanathan,G.M.Revisiting L´e vyﬂightsearch patterns of wandering albatrosses,bumblebees and deer.Nature449,1044-1049(2007). [15]Sohn,T.,Varshavsky,A.,LaMarca,A.,Chen,M.Y.,Choudhury,T.,Smith,I.,Consolvo,S.,High-tower,J.,Griswold,W.G.&de Lara,E.Lecture Notes in Computer Sciences:Proc.8th International Conference UbiComp2006.(Springer,Berlin,2006).[16]Onnela,J.-P.,Saram¨a ki,J.,Hyv¨o nen,J.,Szab´o,G.,Lazer,D.,Kaski,K.,Kert´e sz,K.&Barab´a si A.L.Structure and tie strengths in mobile communication networks.Proceedings of the National Academy of Sciences of the United States of America104,7332-7336(2007).[17]Gonz´a lez,M.C.&Barab´a si,plex networks:From data to models.Nature Physics3,224-225(2007).[18]Palla,G.,Barab´a si,A.-L.&Vicsek,T.Quantifying social group evolution.Nature446,664-667(2007).[19]Hidalgo C.A.&Rodriguez-Sickert C.The dynamics of a mobile phone network.Physica A387,3017-30224.[20]Barab´a si,A.-L.The origin of bursts and heavy tails in human dynamics.Nature435,207-211(2005).[21]Hughes,B.D.Random Walks and Random Environments.(Oxford University Press,USA,1995).[22]Redner,S.A Guide to First-Passage Processes.(Cambridge University Press,UK,2001).[23]Schlich,R.&Axhausen,K.W.Habitual travel behaviour:Evidence from a six-week travel diary.Transportation30,13-36(2003).[24]Eagle,N.&Pentland,A.Eigenbehaviours:Identifying Structure in Routine.submitted to BehavioralEcology and Sociobiology(2007).[25]Yook,S.-H.,Jeong,H.&Barab´a si A.L.Modeling the Internet’s large-scale topology.Proceedings ofthe Nat’l Academy of Sciences99,13382-13386(2002).[26]Caldarelli,G.Scale-Free Networks:Complex Webs in Nature and Technology.(Oxford UniversityPress,USA,2007).[27]Dorogovtsev,S.N.&Mendes,J.F.F.Evolution of Networks:From Biological Nets to the Internet andWWW.(Oxford University Press,USA,2003).[28]Song C.M.,Havlin S.&Makse H.A.Self-similarity of complex networks.Nature433,392-395(2005).[29]Gonz´a lez,M.C.,Lind,P.G.&Herrmann,H.J.A system of mobile agents to model social networks.Physical Review Letters96,088702(2006).[30]Cecconi,F.,Marsili,M.,Banavar,J.R.&Maritan,A.Diffusion,peer pressure,and tailed distributions.Physical Review Letters89,088102(2002).FIG.1:Basic human mobility patterns.a,Week-long trajectory of40mobile phone users indicate that most individuals travel only over short distances,but a few regularly move over hundreds of kilometers. Panel b,displays the detailed trajectory of a single user.The different phone towers are shown as green dots,and the V oronoi lattice in grey marks the approximate reception area of each tower.The dataset studied by us records only the identity of the closest tower to a mobile user,thus we can not identify the position of a user within a V oronoi cell.The trajectory of the user shown in b is constructed from186 two hourly reports,during which the user visited a total of12different locations(tower vicinities).Among these,the user is found96and67occasions in the two most preferred locations,the frequency of visits for each location being shown as a vertical bar.The circle represents the radius of gyration centered in the trajectory’s center of mass.c,Probability density function P(∆r)of travel distances obtained for the two studied datasets D1and D2.The solid line indicates a truncated power law whose parameters are provided in the text(see Eq.1).d,The distribution P(r g)of the radius of gyration measured for the users, where r g(T)was measured after T=6months of observation.The solid line represent a similar truncated power lawﬁt(see Eq.2).The dotted,dashed and dot-dashed curves show P(r g)obtained from the standard null models(RW,LF and T LF),where for the T LF we used the same step size distribution as the onemeasured for the mobile phone users.FIG.2:The bounded nature of human trajectories.a,Radius of gyration, r g(t) vs time for mobile phone users separated in three groups according to theirﬁnal r g(T),where T=6months.The black curves correspond to the analytical predictions for the random walk models,increasing in time as r g(t) |LF,T LF∼t3/2+β(solid),and r g(t) |RW∼t0.5(dotted).The dashed curves corresponding to a logarithmicﬁt of the form A+B ln(t),where A and B depend on r g.b,Probability density function of individual travel distances P(∆r|r g)for users with r g=4,10,40,100and200km.As the inset shows,each group displays a quite different P(∆r|r g)distribution.After rescaling the distance and the distribution with r g(main panel),the different curves collapse.The solid line(power law)is shown as a guide to the eye.c,Return probability distribution,F pt(t).The prominent peaks capture the tendency of humans to regularly return to the locations they visited before,in contrast with the smooth asymptotic behavior∼1/(t ln(t)2)(solid line)predicted for random walks.d,A Zipf plot showing the frequency of visiting different locations.The symbols correspond to users that have been observed to visit n L=5,10,30,and50different locations.Denoting with(L)the rank of the location listed in the order of the visit frequency,the data is well approximated by R(L)∼L−1. The inset is the same plot in linear scale,illustrating that40%of the time individuals are found at theirﬁrsttwo preferred locations.FIG.3:The shape of human trajectories.a,The probability density functionΦ(x,y)ofﬁnding a mobile phone user in a location(x,y)in the user’s intrinsic reference frame(see SM for details).The three plots, from left to right,were generated for10,000users with:r g≤3,20<r g≤30and r g>100km.The trajectories become more anisotropic as r g increases.b,After scaling each position withσx andσy theresulting˜Φ(x/σx,y/σy)has approximately the same shape for each group.c,The change in the shape of Φ(x,y)can be quantiﬁed calculating the isotropy ratio S≡σy/σx as a function of r g,which decreases as S∼r−0.12(solid line).Error bars represent the standard error.d,˜Φ(x/σx,0)representing the x-axis cross gsection of the rescaled distribution˜Φ(x/σx,y/σy)shown in b.。

CFTR与囊性纤维化

第26卷第2期2006年4月国际病理科学与临床杂志I nternati onal Journal of Pathol ogy and ClinicalMedicineVol.26　No.2Ap r.　2006 CFT R与囊性纤维化王　瑞,李学军(北京大学基础医学院药理系,北京100083)[摘要]　囊性纤维化跨膜传导调节因子(CFTR)是一种c AMP激活的ATP门控性氯离子通道,表达于气道,消化道和生殖道上皮细胞的顶部质膜中。

囊性纤维化(CF)是白人中最常见的遗传性疾病之一,由CFT R基因突变造成。

对CFTR基因的破译使人们进一步了解CF的发病机制,并为该疾病的诊断提供了新的线索。

[关键词]　囊性纤维化;　囊性纤维跨膜电导调节因子;　ATP结合盒;　△F508;　基因疗法;　药物治疗[中图分类号]　Q71 [文献标识码]　A [文章编号]　167322588(2006)022*******CFTR and cysti c f i brosisWANG Rui,L I Xue2jun(D epart m ent of Phar m acology,School of B asic M edical Sciences,Peking U niversity,B eijing100083,China)[Abstract]　The cystic fibr osis trans me mbrane conductance regulat or(CFT R)is a c AMP2activa2 ted and ATP2gated Cl-channel exp ressed in the ap ical p las ma me mbrane of ep ithelial cells in the air2 ways,digestive and rep r oductive tracts.Cystic Fibr osis(CF),caused by mutati ons in the CFT R gene,is one of the most common inherited dis orders of white populati ons.The identificati on of the CF gene led us t o a further understanding of the CFT R structure and functi on,the mutati onal basis as well as the com2 p lexity of the disease.[Key words]　cystic fibr osis(CF);　cystic fibr osis trans me mbrane conductance regulat or (CFTR);　ATP binding cassette(ABC);　deltaF508;　gene therapy;　drug treat m ent[In t J Pa thol C lin M ed,2006,26(2):0142204] 囊性纤维化(cystic fibr osis,CF)是一种致命的常染色体隐性疾病,主要临床症状为慢性梗阻性肺部病变,是白人中最常见的遗传性疾病之一。

R包gradientForest使用说明书

Example analysis of biodiversity survey data with R packagegradientForestC.Roland Pitcher,Nick Ellis,Stephen J.SmithMarch21,2011Contents1Introduction1 2Gradient Forest basics22.1Data (2)2.2gradientForest analysis (2)2.3gradientForest plots (3)3Gradient Forest predictions83.1Transforming predictors (9)3.2Biplot of the biological space (9)3.3Mapping in geographic space (12)3.4A clustered version (12)4Session information15 References16 1IntroductionR package gradientForest[Ellis et˜al.,2010]developsﬂexible non-parametric functions to quantify multi-species compositional turnover along environmental gradients.Theﬂexibility comes from the method’s origins in Random Forests[Breiman,2001];speciﬁcally R package randomForest[Liaw and Wiener,2002].This document provides an example to demonstrate the use of gradientForest for ecological analysis of biodiversity survey data.A multi-regional application is provided by[Pitcher et˜al.,2010].The document assumes some familiarity both with R and with community analysis.The example has some analogues with constrained ordi-nation methods and with Generalized Dissimilarity Modelling[Ferrier et˜al.,2007],which are both complementary.Package randomForest includes functions for plotting non-linear responses in compositional along environmental gradients,and for using these responses to transform en-vironmental data layers to biological scales.The transformed multi-dimensional biological space can be represented as a biplot and can be mapped in geographic space.This example demonstrates typical scripts for a gradient forest analysis and provides in the package a sites-by-species(row-by-columns)matrix and a matching sites-by-environment(row-by-columns)dataframe.The number of rows and their order must match between these two data objects.The data should not include NAs.It is assumed that users will be familiar with the data-processing steps necessary to produce such data objects.12Gradient Forest basics2.1DataThe example data provided in package gradientForest are real species data from a cross-shelf survey in the far northern Great Barrier Reef of northeast Australia[Poiner et˜al.,1998, Burridge et˜al.,2006].Of>1,000species observed,a subset of110are included from197of the sites sampled.The environmental data include28predictors,either measured at each site or attributed to each site by interpolation[Pitcher et˜al.,2002].>require(gradientForest)>load("GZ.sps.mat.Rdata")>dim(Sp_mat)[1]197110>load("GZ.phys.site.Rdata")>dim(Phys_site)[1]197282.2gradientForest analysisThe function gradientForest is a wrapper function that calls extendedForest,a modiﬁed version of randomForest,and collates its output across all the species in the data matrix. The key modiﬁcation in extendedForest extracts the numerous tree split values along each predictor gradient and their associatedﬁt improvement,for each predictor in each tree,for the forests and returns that information to gradientForest.Like randomForest,extendedForest assesses the importance of each variable for predic-tion accuracy;information that is further collated and processed by gradientForest.Often, predictor variables are correlated however.The standard approach in random forests assesses marginal importance of predictor by randomly permuting each predictor in turn,across all sites in the dataset,and calculating the degradation prediction performance of each tree.Package extendedForest can account for correlated predictors by implementing conditional permutation [Ellis et˜al.,2010],following the strategy outlined by Strobl et˜al.[2008].In conditional per-mutation,the predictor to be assessed is permuted only within blocks of the dataset deﬁned by splits in the given tree on any other predictors correlated above a certain threshold(e.g. r=0.5)and up to a maximum number of splits set by the maxLevel option(if required).>nSites<-dim(Sp_mat)[1]>nSpecs<-dim(Sp_mat)[2]>lev<-floor(log2(nSites*0.368/2))>lev[1]5The gradientForest may take several minutes to run.Other options that can be set include the number of trees typically500,whether the splits should be compact into bins(advising to prevent memory problems for large datasets)and the number of bins,and the correlation threshold for conditional permutation.The summary shows the number of species with positive R2ie.those species that could be predicted to any extent by the available predictor.The returned object is a list containing the data,predictor importances,species R2’s and other information described in the html help pages under Value.>gf<-gradientForest(cbind(Phys_site,Sp_mat),+predictor.vars=colnames(Phys_site),response.vars=colnames(Sp_mat),+ntree=500,transform=NULL,compact=T,+nbin=201,maxLevel=lev,corr.threshold=0.5)>gfA forest of500regression trees for each of90speciesCall:gradientForest(data=cbind(Phys_site,Sp_mat),predictor.vars=colnames(Phys_site), response.vars=colnames(Sp_mat),ntree=500,transform=NULL,maxLevel=lev,corr.threshold=0.5,compact=T,nbin=201)Important variables:[1]BSTRESS MUD S_AV Si_AV CHLA_AV>names(gf)[1]"X""Y""result"[4]"overall.imp""overall.imp2""ntree"[7]"imp.rsq""species.pos.rsq""ranForest.type"[10]"res""res.u""dens"[13]"call"2.3gradientForest plotsSeveral types of plots are available for the gradientForest object.Theﬁrst is the predictor overall importance plot.This show the mean accuracy importance and the mean importance weighted by species R2.in this example,both are conditional importance.Seabed stress and sediment mud fraction are clearly the most important variables across these89species.>plot(gf,plot.type="O")PO4_SRASPECT SLOPE O2_SR O2_AV PO4_AV CHLA_SR S_SR K490_SRSi_SR SST_SR BIR_AV BIR_SR GRAVEL BA THY NO3_SR T_AV T_SR SST_AV CRBNT NO3_AV Si_AV SAND K490_AV CHLA_AV S_AV MUD BSTRESSAccuracy importance0.000.040.08PO4_SRASPECT O2_SR SLOPE O2_AV CHLA_SR PO4_AV Si_SR S_SR BIR_AV BIR_SR K490_SR SST_SR NO3_SR BA THY T_SR GRAVEL SST_AV NO3_AV T_AV CRBNT K490_AV SAND CHLA_AVSi_AV S_AV MUD BSTRESSR 2weighted importance0.0000.0100.020The predictor gradient plots are best presented in order of importance;in this example the top 25predictors are presented in 5by 5panels.>most_important <-names(importance(gf))[1:25]>par(mgp =c(2,0.75,0))The second plot is the splits density plot (plot.type="S"),which shows binned split impor-tance and location on each gradient (spikes),kernel density of splits (black lines ),of observations (red lines )and of splits standardised by observations density (blue lines ).Each distribution in-tegrates to predictor importance.These show where important changes in the abundance of multiple species are occurring along the gradient;they indicate a composition change rate.Many of the usual plot options can be set in the call.>plot(gf,plot.type ="S",imp.vars =most_important,+leg.posn ="topright",cex.legend =0.4,cex.axis =0.6,+b =0.7,line.ylab =0.9,par.args =list(mgp =c(1.5,+0.5,0),mar =c(3.1,1.5,0.1,1)))0.10.30.50.000.040.080.12BSTRESS02040600e +004e −048e −04MUD 34.835.20.000.040.080.12S_AV 1.5 2.5 3.50.0000.0060.012Si_AV0.51.0 1.52.00.000.020.04CHLA_AV020*******.000000.00015SAND0.060.100.140.00.20.40.6K490_AV4060801000.00000.00100.0020CRBNT27.528.028.529.00.0000.010T_AV0.20.30.40.50.000.050.100.15NO3_AV26.726.927.10.000.040.08SST_AV020*******.000000.00015GRAVEL1.02.03.00.0000.0060.012T_SR−50−40−30−20−100e +003e −046e −04BATHY 0.10.20.30.40.50.000.020.040.06NO3_SR4.2 4.65.0 5.40.0000.0060.012SST_SR 0.050.100.150.200.000.100.20K490_SR0.050.150.250.000.020.04BIR_SR0.00.20.40.60.80.0000.0150.030BIR_AV 1.02.03.00.0000.0020.004S_SR1234560.00000.0010Si_SR 0.140.180.000.10PO4_AV 123450.0000.0040.008CHLA_SR 4.254.35 4.450.0000.0100.020O2_AV 0.0 1.0 2.00.0000.0020.0040.006SLOPED e n s i t yThe third plot is the species cumulative plot (plot.type="C",show.overall=F ),which for each species shows cumulative importance distributions of splits improvement scaled by R 2weighted importance,and standardised by density of observations.These show cumulative change in abundance of individual species,where changes occur on the gradient,and the species changing most on each gradient.Again many of the usual plot options can be set in the call;in this example the legend identiﬁes the top 5most responsive species for each predictor >plot(gf,plot.type ="C",imp.vars =most_important,+show.overall =F,legend =T,leg.posn ="topleft",+leg.nspecies =5,b =0.7,cex.legend =0.4,+cex.axis =0.6,line.ylab =0.9,par.args =list(mgp =c(1.5,+0.5,0),mar =c(2.5,1,0.1,0.5),omi =c(0,+0.3,0,0)))0.10.20.30.40.50.000.100.20BSTRESS01030500.000.020.040.06MUD34.835.035.235.40.000.040.08S_AV1.52.53.50.000.040.080.12Si_AV0.5 1.0 1.5 2.00.000.020.04CHLA_AV0204060801000.000.040.08SAND0.060.100.140.000.020.040.06K490_AV4060801000.000.020.04CRBNT27.528.028.529.00.000.020.040.06T_AV0.20.30.40.50.000.020.04NO3_AV26.726.826.927.027.10.000.020.04SST_AV0204060801000.000.020.04GRAVEL1.0 1.52.0 2.53.0 3.50.000.020.04T_SR−50−40−30−20−100.000.020.04BATHY0.10.20.30.40.50.000.020.04NO3_SR4.2 4.65.0 5.40.0000.0100.020SST_SR0.050.100.150.200.000.020.040.06K490_SR0.050.150.250.0000.0100.020BIR_SR0.00.20.40.60.80.0000.0100.020BIR_AV1.02.03.00.000.020.04S_SR1234560.000.020.04Si_SR 0.140.160.180.200.0000.0150.030PO4_AV 123450.0000.0100.0200.030CHLA_SR 4.254.354.450.0000.0040.0080.012O2_AV 0.00.5 1.0 1.5 2.0 2.50.0000.0100.0200.030SLOPEC u m u l a t i v e i m p o r t a n c eThe fourth plot is the predictor cumulative plot (plot.type="C",show.species=F ),which for each predictor shows cumulative importance distributions of splits improvement scaled by R 2weighted importance,and standardised by density of observations,averaged over all species.These show cumulative change in overall composition of the community,where changes occur on the gradient.Again many of the usual plot options can be set in the call;in this example com-mon.scale=T ensures that plots for all predictors have the same y-scale as the most important predictor.>plot(gf,plot.type ="C",imp.vars =most_important,+show.species =F,common.scale =T,cex.axis =0.6,+b =0.7,line.ylab =0.9,par.args =list(mgp =c(1.5,+0.5,0),mar =c(2.5,1,0.1,0.5),omi =c(0,+0.3,0,0)))0.10.20.30.40.50.0000.0100.020BSTRESS01030500.0000.0100.020MUD34.835.035.235.40.0000.0100.020S_AV1.52.53.50.0000.0100.020Si_AV0.51.0 1.52.00.0000.0100.020CHLA_AV0204060801000.0000.0100.020SAND0.060.100.140.0000.0100.020K490_AV4060801000.0000.0100.020CRBNT27.528.028.529.00.0000.0100.020T_AV0.20.30.40.50.0000.0100.020NO3_AV26.726.826.927.027.10.0000.0100.020SST_AV0204060801000.0000.0100.020GRAVEL1.0 1.52.0 2.53.0 3.50.0000.0100.020T_SR−50−40−30−20−100.0000.0100.020BATHY0.10.20.30.40.50.0000.0100.020NO3_SR4.2 4.65.0 5.40.0000.0100.020SST_SR0.050.100.150.200.0000.0100.020K490_SR0.050.150.250.0000.0100.020BIR_SR0.00.20.40.60.80.0000.0100.020BIR_AV1.02.03.00.0000.0100.020S_SR1234560.0000.0100.020Si_SR 0.140.160.180.200.0000.0100.020PO4_AV 123450.0000.0100.020CHLA_SR 4.25 4.35 4.450.0000.0100.020O2_AV0.00.5 1.0 1.5 2.0 2.50.0000.0100.020SLOPEC u m u l a t i v e i m p o r t a n c eThe ﬁfth plot shows the R 2measure of the ﬁt of the random forest model for each species,ordered in various ways.>plot(gf,plot.type ="P",s =F,horizontal =F,+cex.axis =1,bels =0.7,line =2.5)qq q q qq q q qq q q q qqq q q q qq q q qq q q q q q q q qq q q q q q q q q q q q q q q q q q qq q q q q q q q q qq q q q q qq q q q q q q q qq q q q q q q q q q q q q1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889900.00.10.20.30.40.50.6Species performance rankR 2Overall performance of random forests over speciesSeveral other alternative formats of the R 2ﬁt performance plot are available,e.g.:>plot(gf,plot.type ="P",s =T,horizontal =F,+cex.axis =1,bels =0.7,line =2.5)>plot(gf,plot.type ="P",s =F,horizontal =T,+cex.axis =1,bels =0.6,line =2.5)>plot(gf,plot.type ="P",s =T,horizontal =T,+cex.axis =1,bels =0.6,line =2.5)3Gradient Forest predictionsIn addition to examining compositional change along environmental gradients,the predictor cumulative functions can also be used to transform grid data layers of environmental variables to a common biological importance scale.This transformation of the multi-dimensional en-vironment space is to a biological space in which coordinate position represents compositionassociated with the predictors.These inferred compositional patterns can be mapped in bio-logical space and geographic space in a manner analogous to ordination,but takes into account the non-linear and sometimes threshold changes that occur along gradients.3.1Transforming predictorsThe example provided includes gridded environmental variables for a roughly10,000km2area of the far northern Great Barrier Reef where the biological surveys were conducted.The data include North and East coordinates plus28predictors at8,682grid cells.The grid data must include the same predictors with the same names as sites included in the gradientForest call.>load("GZ.phys.grid.Rdata")>dim(Phys_grid)[1]868230>names(Phys_grid)[1]"NORTH""EAST""BATHY""SLOPE""ASPECT" [6]"BSTRESS""CRBNT""GRAVEL""SAND""MUD"[11]"NO3_AV""NO3_SR""PO4_AV""PO4_SR""O2_AV"[16]"O2_SR""S_AV""S_SR""T_AV""T_SR"[21]"Si_AV""Si_SR""CHLA_AV""CHLA_SR""K490_AV"[26]"K490_SR""SST_AV""SST_SR""BIR_AV""BIR_SR"The grid variables are transformed using the gradientForest predict function.>imp.vars<-names(importance(gf))>Trns_grid<-cbind(Phys_grid[,c("EAST","NORTH")],+predict(gf,Phys_grid[,imp.vars]))It is useful to also transform the site environmental predictors,which are available from gf$X.>Trns_site<-predict(gf)3.2Biplot of the biological spaceThe multi-dimensional biological space can most eﬀectively be represented by taking the prin-ciple components of the transformed grid and presenting theﬁrst two dimensions in a biplot. It must be acknowledged that while most variation in patterns is captured by theﬁrst dimen-sions,additional compositional pattern contained in the higher dimensions is not shown.A user deﬁned RGB colour palette is set up based on theﬁrst3dimensions.>PCs<-prcomp(Trns_grid[,imp.vars])>a1<-PCs$x[,1]>a2<-PCs$x[,2]>a3<-PCs$x[,3]>r<-a1+a2>g<--a2>b<-a3+a2-a1>r<-(r-min(r))/(max(r)-min(r))*255>g<-(g-min(g))/(max(g)-min(g))*255>b<-(b-min(b))/(max(b)-min(b))*255The environmental variables may be shown as vectors,perhaps limited to the most important predictors—in this example,variables to show as vectors are selected.>nvs<-dim(PCs$rotation)[1]>vec<-c("BSTRESS","MUD","SST_AV","T_AV","CHLA_AV", +"SAND","CRBNT","GRAVEL")>lv<-length(vec)>vind<-rownames(PCs$rotation)%in%vec>scal<-40>xrng<-range(PCs$x[,1],PCs$rotation[,1]/scal)*+ 1.1>yrng<-range(PCs$x[,2],PCs$rotation[,2]/scal)*+ 1.1>plot((PCs$x[,1:2]),xlim=xrng,ylim=yrng,+pch=".",cex=4,col=rgb(r,g,b,max=255),+asp=1)>points(PCs$rotation[!vind,1:2]/scal,pch="+")>arrows(rep(0,lv),rep(0,lv),PCs$rotation[vec,+1]/scal,PCs$rotation[vec,2]/scal,length=0.0625)>jit<-0.0015>text(PCs$rotation[vec,1]/scal+jit*sign(PCs$rotation[vec,+1]),PCs$rotation[vec,2]/scal+jit*sign(PCs$rotation[vec,+2]),labels=vec)Diﬀerent coordinate positions in the biplot represent diﬀering compositions,as associated with the predictors.Further information may be added to the biplot including the location of sites in biological space,the weight mean location of species,and selected species may be identiﬁed interactively.−0.02−0.010.000.010.02−0.02−0.010.000.010.02PC1P C 2>PCsites <-predict(PCs,Trns_site[,imp.vars])>points(PCsites[,1:2])>SpsWtd <-sweep(gf$Y,2,apply(gf$Y,2,min),+"-")>SpsWtdPCs <-(t(SpsWtd)%*%(PCsites[,1:2]))/colSums(SpsWtd)>points(SpsWtdPCs,col ="red",pch ="+")If required,the abundance of any given species may be plotted on the biplot.For example the ﬁrst species from gf$Y =A1010102,an alga from the family Caulerpaceae that appears to prefer carbonate gravelly sand area with moderate bedstress and lower temperature.>sp <-colnames(SpsWtd)[1]>points(PCsites[,1:2],col ="blue",cex =SpsWtd[,+sp]/2)Alternatively,speciﬁcally named examples could be plotted:e.g.E4030373a Fungiid coral;M2020101a Strombid mollusc;or S1010671a Perophorid ascidian to name a few.3.3Mapping in geographic spaceThe biplot and the colour palette in the previous section can be used as a key to visualise com-positional patterns mapped in geographic space.The following map plots predicted PC scores in geographic coordinates,using the same colour palette as above,and represents continuous changes in inferred compositional patterns associated with the predictors.>plot(Trns_grid[,c("EAST","NORTH")],pch =".",+cex =3,asp =1,col =rgb(r,g,b,max =255))−0.6−0.4−0.20.00.20.40.6−0.6−0.4−0.20.00.20.4EASTN O R T H3.4A clustered versionSome applications may require a hard clustered output,representing inferred assemblages,rather than a continuous representation of biodiversity composition.The following example uses clara to make 8clusters.This is a fast claustering algorithm suitable for large data sets.The medoids are labelled and the colour key takes the value for each medoid.Other clustering methods maybe used(for example,pam would take several minutes)as alternatives,and their various cluster diagnostics may provide a guide to the appropriate numbers of clusters.>require(cluster)>ncl<-8>clPCs<-clara(PCs$x,ncl,sampsize=1000)>medcolR<-r[clPCs$i.med]>medcolG<-g[clPCs$i.med]>medcolB<-b[clPCs$i.med]>plot((PCs$x[,1:2]),xlim=xrng,ylim=yrng,+pch=".",cex=4,col=rgb(medcolR[clPCs$clustering],+medcolG[clPCs$clustering],medcolB[clPCs$clustering],+max=255),asp=1)>points(PCs$rotation[!vind,1:2]/scal,pch="+")>arrows(rep(0,lv),rep(0,lv),PCs$rotation[vec,+1]/scal,PCs$rotation[vec,2]/scal,length=0.0625)>text(PCs$rotation[vec,1]/scal+jit*sign(PCs$rotation[vec,+1]),PCs$rotation[vec,2]/scal+jit*sign(PCs$rotation[vec,+2]),labels=vec)>text(clPCs$medoids[,1:2],labels=seq(1,ncl))>legend("bottomleft",as.character(seq(1,ncl)),+pch=15,cex=1,col=rgb(medcolR,medcolG,+medcolB,max=255))−0.02−0.010.000.010.02−0.02−0.010.000.010.02PC1P C 2>plot(Trns_grid[,c("EAST","NORTH")],pch =".",+cex =3,asp =1,col =rgb(medcolR[clPCs$clustering],+medcolG[clPCs$clustering],medcolB[clPCs$clustering],+max =255))>points(Trns_grid[clPCs$i.med,c("EAST","NORTH")],+pch =as.character(seq(1,ncl)))>legend("bottomleft",as.character(seq(1,ncl)),+pch =15,cex =1,col =rgb(medcolR,medcolG,+medcolB,max =255))−0.6−0.4−0.20.00.20.40.6−0.6−0.4−0.20.00.20.4EASTN O R T H4Session informationThe simulation and output in this document were generated in the following computing envi-ronment.•R version 2.12.2Patched (2011-03-18r54866),x86_64-unknown-linux-gnu•Locale:LC_CTYPE=en_US.UTF-8,LC_NUMERIC=C ,LC_TIME=en_US.UTF-8,LC_COLLATE=C ,LC_MONETARY=C ,LC_MESSAGES=en_US.UTF-8,LC_PAPER=en_US.UTF-8,LC_NAME=C ,LC_ADDRESS=C ,LC_TELEPHONE=C ,LC_MEASUREMENT=en_US.UTF-8,LC_IDENTIFICATION=C•Base packages:base,datasets,grDevices,graphics,methods,stats,utils •Other packages:cluster˜1.13.3,extendedForest˜1.4,gradientForest˜0.1-11,lattice˜0.19-17•Loaded via a namespace(and not attached):grid˜2.12.2,tools˜2.12.2ReferencesL.˜Breiman.Random Forests.Machine Learning,45(1):5–32,2001.C.Y.Burridge,C.R.Pitcher,B.J.Hill,T.J.Wassenberg,and I.R.Poiner.A comparison of demersal communities in an area closed to trawling with those in adjacent areas open to trawling:a study in the Great Barrier Reef Marine Park,Australia.Fisheries Research,79: 64–74,2006.N.˜Ellis,S.J.Smith,and C.R.Pitcher.Gradient forests:calculating importance gradients on physical predictors.submitted manuscript.2010.S.˜Ferrier,G.˜Manion,J.˜Elith,and K.˜ing generalized dissimilarity modelling to analyse and predict patterns of beta diversity in regional biodiversity assessment.Diversity and Distributions,13(3):252–264,2007.Andy Liaw and Matthew Wiener.Classiﬁcation and regression by randomforest.R News,2(3): 18–22,2002.URL /doc/Rnews/.C.R.Pitcher,W.˜Venables,N.˜Ellis,I.˜McLeod,M.˜Cappo, F.˜Pantus,M.˜Austin, P.˜Doherty,and N.˜Gribble.Gbr seabed biodiversity mapping project:Phase1 report to crc-reef.Technical report,CSIRO/AIMS/QDPI Report,2002.URL .au/resprogram/programC/seabed/Seabedphase1rpt.htm.C.R.Pitcher,P.˜Lawton,N.˜Ellis,S.J.Smith,L.S.Incze,C-L.Wei,M.E.Greenlaw,N.H.Wolﬀ, J.˜Sameoto,and P.V.R.Snelgrove.The role of physical environmental variables in shaping patterns of biodiversity composition in seabed assemblages.submitted manuscript.2010.IR˜Poiner,J.˜Glaister,CR˜Pitcher, C.˜Burridge,T.˜Wassenberg,N.˜Gribble, B.˜Hill,SJM Blaber,DA˜Milton, D.˜Brewer,et˜al.The environmental ef-fects of prawn trawling in the far northern section of the Great Barrier Reef Marine Park:1991–1996.Final Report to GBRMPA and FRDC,1998.URL http://www.publish.csiro.au/books/bookpage.cfm?PID=2419.C.˜Strobl,A.L.Boulesteix,T.˜Kneib,T.˜Augustin,and A.˜Zeileis.Conditional variable im-portance for random forests.BMC Bioinformatics,9(1):307,2008.。

1999.Multilevel Hypergraph Partitioning__Applications in VLSI Domain

Multilevel Hypergraph Partitioning:Applications in VLSI DomainGeorge Karypis,Rajat Aggarwal,Vipin Kumar,Senior Member,IEEE,and Shashi Shekhar,Senior Member,IEEE Abstract—In this paper,we present a new hypergraph-partitioning algorithm that is based on the multilevel paradigm.In the multilevel paradigm,a sequence of successivelycoarser hypergraphs is constructed.A bisection of the smallesthypergraph is computed and it is used to obtain a bisection of theoriginal hypergraph by successively projecting and reﬁning thebisection to the next levelﬁner hypergraph.We have developednew hypergraph coarsening strategies within the multilevelframework.We evaluate their performance both in terms of thesize of the hyperedge cut on the bisection,as well as on the runtime for a number of very large scale integration circuits.Ourexperiments show that our multilevel hypergraph-partitioningalgorithm produces high-quality partitioning in a relatively smallamount of time.The quality of the partitionings produced by ourscheme are on the average6%–23%better than those producedby other state-of-the-art schemes.Furthermore,our partitioningalgorithm is signiﬁcantly faster,often requiring4–10times lesstime than that required by the other schemes.Our multilevelhypergraph-partitioning algorithm scales very well for largehypergraphs.Hypergraphs with over100000vertices can bebisected in a few minutes on today’s workstations.Also,on thelarge hypergraphs,our scheme outperforms other schemes(inhyperedge cut)quite consistently with larger margins(9%–30%).Index Terms—Circuit partitioning,hypergraph partitioning,multilevel algorithms.I.I NTRODUCTIONH YPERGRAPH partitioning is an important problem withextensive application to many areas,including very largescale integration(VLSI)design[1],efﬁcient storage of largedatabases on disks[2],and data mining[3].The problemis to partition the vertices of a hypergraphintois deﬁnedas a set ofvertices[4],and the size ofa hyperedge is the cardinality of this subset.Manuscript received April29,1997;revised March23,1998.This workwas supported under IBM Partnership Award NSF CCR-9423082,by theArmy Research Ofﬁce under Contract DA/DAAH04-95-1-0538,and by theArmy High Performance Computing Research Center,the Department of theArmy,Army Research Laboratory Cooperative Agreement DAAH04-95-2-0003/Contract DAAH04-95-C-0008.G.Karypis,V.Kumar,and S.Shekhar are with the Department of ComputerScience and Engineering,Minneapolis,University of Minnesota,Minneapolis,MN55455-0159USA.R.Aggarwal is with the Lattice Semiconductor Corporation,Milpitas,CA95131USA.Publisher Item Identiﬁer S1063-8210(99)00695-2.During the course of VLSI circuit design and synthesis,itis important to be able to divide the system speciﬁcation intoclusters so that the inter-cluster connections are minimized.This step has many applications including design packaging,HDL-based synthesis,design optimization,rapid prototyping,simulation,and testing.In particular,many rapid prototyp-ing systems use partitioning to map a complex circuit ontohundreds of interconnectedﬁeld-programmable gate arrays(FPGA’s).Such partitioning instances are challenging becausethe timing,area,and input/output(I/O)resource utilizationmust satisfy hard device-speciﬁc constraints.For example,ifthe number of signal nets leaving any one of the clustersis greater than the number of signal p-i-n’s available in theFPGA,then this cluster cannot be implemented using a singleFPGA.In this case,the circuit needs to be further partitioned,and thus implemented using multiple FPGA’s.Hypergraphscan be used to naturally represent a VLSI circuit.The verticesof the hypergraph can be used to represent the cells of thecircuit,and the hyperedges can be used to represent the netsconnecting these cells.A high quality hypergraph-partitioningalgorithm greatly affects the feasibility,quality,and cost ofthe resulting system.A.Related WorkThe problem of computing an optimal bisection of a hy-pergraph is at least NP-hard[5].However,because of theimportance of the problem in many application areas,manyheuristic algorithms have been developed.The survey byAlpert and Khang[1]provides a detailed description andcomparison of such various schemes.In a widely used class ofiterative reﬁnement partitioning algorithms,an initial bisectionis computed(often obtained randomly)and then the partitionis reﬁned by repeatedly moving vertices between the twoparts to reduce the hyperedge cut.These algorithms oftenuse the Schweikert–Kernighan heuristic[6](an extension ofthe Kernighan–Lin(KL)heuristic[7]for hypergraphs),or thefaster Fiduccia–Mattheyses(FM)[8]reﬁnement heuristic,toiteratively improve the quality of the partition.In all of thesemethods(sometimes also called KLFM schemes),a vertex ismoved(or a vertex pair is swapped)if it produces the greatestreduction in the edge cuts,which is also called the gain formoving the vertex.The partition produced by these methodsis often poor,especially for larger hypergraphs.Hence,thesealgorithms have been extended in a number of ways[9]–[12].Krishnamurthy[9]tried to introduce intelligence in the tie-breaking process from among the many possible moves withthe same high gain.He used a Look Ahead()algorithm,which looks ahead uptoa move.PROP [11],introduced by Dutt and Deng,used a probabilistic gain computation model for deciding which vertices need to move across the partition line.These schemes tend to enhance the performance of the basic KLFM family of reﬁnement algorithms,at the expense of increased run time.Dutt and Deng [12]proposed two new methods,namely,CLIP and CDIP,for computing the gains of hyperedges that contain more than one node on either side of the partition boundary.CDIP in conjunctionwithand CLIP in conjunction with PROP are two schemes that have shown the best results in their experiments.Another class of hypergraph-partitioning algorithms [13]–[16]performs partitioning in two phases.In the ﬁrst phase,the hypergraph is coarsened to form a small hypergraph,and then the FM algorithm is used to bisect the small hypergraph.In the second phase,these algorithms use the bisection of this contracted hypergraph to obtain a bisection of the original hypergraph.Since FM reﬁnement is done only on the small coarse hypergraph,this step is usually fast,but the overall performance of such a scheme depends upon the quality of the coarsening method.In many schemes,the projected partition is further improved using the FM reﬁnement scheme [15].Recently,a new class of partitioning algorithms was devel-oped [17]–[20]based upon the multilevel paradigm.In these algorithms,a sequence of successively smaller (coarser)graphs is constructed.A bisection of the smallest graph is computed.This bisection is now successively projected to the next-level ﬁner graph and,at each level,an iterative reﬁnement algorithm such as KLFM is used to further improve the bisection.The various phases of multilevel bisection are illustrated in Fig.1.Iterative reﬁnement schemes such as KLFM become quite powerful in this multilevel context for the following reason.First,the movement of a single node across a partition bound-ary in a coarse graph can lead to the movement of a large num-ber of related nodes in the original graph.Second,the reﬁned partitioning projected to the next level serves as an excellent initial partitioning for the KL or FM reﬁnement algorithms.This paradigm was independently studied by Bui and Jones [17]in the context of computing ﬁll-reducing matrix reorder-ing,by Hendrickson and Leland [18]in the context of ﬁnite-element mesh-partitioning,and by Hauck and Borriello (called Optimized KLFM)[20],and by Cong and Smith [19]for hy-pergraph partitioning.Karypis and Kumar extensively studied this paradigm in [21]and [22]for the partitioning of graphs.They presented new graph coarsening schemes for which even a good bisection of the coarsest graph is a pretty good bisec-tion of the original graph.This makes the overall multilevel paradigm even more robust.Furthermore,it allows the use of simpliﬁed variants of KLFM reﬁnement schemes during the uncoarsening phase,which signiﬁcantly speeds up the reﬁne-ment process without compromising overall quality.METIS [21],a multilevel graph partitioning algorithm based upon this work,routinely ﬁnds substantially better bisections and is often two orders of magnitude faster than the hitherto state-of-the-art spectral-based bisection techniques [23],[24]for graphs.The improved coarsening schemes of METIS work only for graphs and are not directly applicable to hypergraphs.IftheFig.1.The various phases of the multilevel graph bisection.During the coarsening phase,the size of the graph is successively decreased;during the initial partitioning phase,a bisection of the smaller graph is computed,and during the uncoarsening and reﬁnement phase,the bisection is successively reﬁned as it is projected to the larger graphs.During the uncoarsening and reﬁnement phase,the dashed lines indicate projected partitionings and dark solid lines indicate partitionings that were produced after reﬁnement.G 0is the given graph,which is the ﬁnest graph.G i +1is the next level coarser graph of G i ,and vice versa,G i is the next level ﬁner graph of G i +1.G 4is the coarsest graph.hypergraph is ﬁrst converted into a graph (by replacing each hyperedge by a set of regular edges),then METIS [21]can be used to compute a partitioning of this graph.This technique was investigated by Alpert and Khang [25]in their algorithm called GMetis.They converted hypergraphs to graphs by simply replacing each hyperedge with a clique,and then they dropped many edges from each clique randomly.They used METIS to compute a partitioning of each such random graph and then selected the best of these partitionings.Their results show that reasonably good partitionings can be obtained in a reasonable amount of time for a variety of benchmark problems.In particular,the performance of their resulting scheme is comparable to other state-of-the art schemes such as PARABOLI [26],PROP [11],and the multilevel hypergraph partitioner from Hauck and Borriello [20].The conversion of a hypergraph into a graph by replacing each hyperedge with a clique does not result in an equivalent representation since high-quality partitionings of the resulting graph do not necessarily lead to high-quality partitionings of the hypergraph.The standard hyperedge-to-edge conversion [27]assigns a uniform weightofisthe of the hyperedge,i.e.,thenumber of vertices in the hyperedge.However,the fundamen-tal problem associated with replacing a hyperedge by its clique is that there exists no scheme to assign weight to the edges of the clique that can correctly capture the cost of cutting this hyperedge [28].This hinders the partitioning reﬁnement algo-rithm since vertices are moved between partitions depending on how much this reduces the number of edges they cut in the converted graph,whereas the real objective is to minimize the number of hyperedges cut in the original hypergraph.Furthermore,the hyperedge-to-clique conversion destroys the natural sparsity of the hypergraph,signiﬁcantly increasing theKARYPIS et al.:MULTILEVEL HYPERGRAPH PARTITIONING:APPLICATIONS IN VLSI DOMAIN 71run time of the partitioning algorithm.Alpert and Khang [25]solved this problem by dropping many edges of the clique randomly,but this makes the graph representation even less accurate.A better approach is to develop coarsening and reﬁnement schemes that operate directly on the hypergraph.Note that the multilevel scheme by Hauck and Borriello [20]operates directly on hypergraphs and,thus,is able to perform accurate reﬁnement during the uncoarsening phase.However,all coarsening schemes studied in [20]are edge-oriented;i.e.,they only merge pairs of nodes to construct coarser graphs.Hence,despite a powerful reﬁnement scheme (FM with theuse oflook-ahead)during the uncoarsening phase,their performance is only as good as that of GMetis [25].B.Our ContributionsIn this paper,we present a multilevel hypergraph-partitioning algorithm hMETIS that operates directly on the hypergraphs.A key contribution of our work is the development of new hypergraph coarsening schemes that allow the multilevel paradigm to provide high-quality partitions quite consistently.The use of these powerful coarsening schemes also allows the reﬁnement process to be simpliﬁed considerably (even beyond plain FM reﬁnement),making the multilevel scheme quite fast.We investigate various algorithms for the coarsening and uncoarsening phases which operate on the hypergraphs without converting them into graphs.We have also developed new multiphase reﬁnement schemes(-cycles)based on the multilevel paradigm.These schemes take an initial partition as input and try to improve them using the multilevel scheme.These multiphase schemes further reduce the run times,as well as improve the solution quality.We evaluate their performance both in terms of the size of the hyperedge cut on the bisection,as well as on run time on a number of VLSI circuits.Our experiments show that our multilevel hypergraph-partitioning algorithm produces high-quality partitioning in a relatively small amount of time.The quality of the partitionings produced by our scheme are on the average 6%–23%better than those produced by other state-of-the-art schemes [11],[12],[25],[26],[29].The difference in quality over other schemes becomes even greater for larger hypergraphs.Furthermore,our partitioning algorithm is signiﬁcantly faster,often requiring 4–10times less time than that required by the other schemes.For many circuits in the well-known ACM/SIGDA benchmark set [30],our scheme is able to ﬁnd better partitionings than those reported in the literature for any other hypergraph-partitioning algorithm.The remainder of this paper is organized as follows.Section II describes the different algorithms used in the three phases of our multilevel hypergraph-partitioning algorithm.Section III describes a new partitioning reﬁnement algorithm based on the multilevel paradigm.Section IV compares the results produced by our algorithm to those produced by earlier hypergraph-partitioning algorithms.II.M ULTILEVEL H YPERGRAPH B ISECTIONWe now present the framework of hMETIS ,in which the coarsening and reﬁnement scheme work directly with hyper-edges without using the clique representation to transform them into edges.We have developed new algorithms for both the phases,which,in conjunction,are capable of delivering very good quality solutions.A.Coarsening PhaseDuring the coarsening phase,a sequence of successively smaller hypergraphs are constructed.As in the case of mul-tilevel graph bisection,the purpose of coarsening is to create a small hypergraph,such that a good bisection of the small hypergraph is not signiﬁcantly worse than the bisection di-rectly obtained for the original hypergraph.In addition to that,hypergraph coarsening also helps in successively reducing the sizes of the hyperedges.That is,after several levels of coarsening,large hyperedges are contracted to hyperedges that connect just a few vertices.This is particularly helpful,since reﬁnement heuristics based on the KLFM family of algorithms [6]–[8]are very effective in reﬁning small hyperedges,but are quite ineffective in reﬁning hyperedges with a large number of vertices belonging to different partitions.Groups of vertices that are merged together to form single vertices in the next-level coarse hypergraph can be selected in different ways.One possibility is to select pairs of vertices with common hyperedges and to merge them together,as illustrated in Fig.2(a).A second possibility is to merge together all the vertices that belong to a hyperedge,as illustrated in Fig.2(b).Finally,a third possibility is to merge together a subset of the vertices belonging to a hyperedge,as illustrated in Fig.2(c).These three different schemes for grouping vertices together for contraction are described below.1)Edge Coarsening (EC):The heavy-edge matching scheme used in the multilevel-graph bisection algorithm can also be used to obtain successively coarser hypergraphs by merging the pairs of vertices connected by many hyperedges.In this EC scheme,a heavy-edge maximal 1matching of the vertices of the hypergraph is computed as follows.The vertices are visited in a random order.For eachvertex are considered,and the one that is connected via the edge with the largest weight is matchedwithandandofsize72IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION(VLSI)SYSTEMS,VOL.7,NO.1,MARCH1999Fig.2.Various ways of matching the vertices in the hypergraph and the coarsening they induce.(a)In edge-coarsening,connected pairs of vertices are matched together.(b)In hyperedge-coarsening,all the vertices belonging to a hyperedge are matched together.(c)In MHEC,we match together all the vertices in a hyperedge,as well as all the groups of vertices belonging to a hyperedge.weight of successively coarser graphs does not decrease very fast.In order to ensure that for every group of vertices that are contracted together,there is a decrease in the hyperedge weight in the coarser graph,each such group of vertices must be connected by a hyperedge.This is the motivation behind the HEC scheme.In this scheme,an independent set of hyperedges is selected and the vertices that belong to individual hyperedges are contracted together.This is implemented as follows.The hyperedges are initially sorted in a nonincreasing hyperedge-weight order and the hyperedges of the same weight are sorted in a nondecreasing hyperedge size order.Then,the hyperedges are visited in that order,and for each hyperedge that connects vertices that have not yet been matched,the vertices are matched together.Thus,this scheme gives preference to the hyperedges that have large weight and those that are of small size.After all of the hyperedges have been visited,the groups of vertices that have been matched are contracted together to form the next level coarser graph.The vertices that are not part of any contracted hyperedges are simply copied to the next level coarser graph.3)Modiﬁed Hyperedge Coarsening(MHEC):The HEC algorithm is able to signiﬁcantly reduce the amount of hyperedge weight that is left exposed in successively coarser graphs.However,during each coarsening phase,a majority of the hyperedges do not get contracted because vertices that belong to them have been contracted via other hyperedges. This leads to two problems.First,the size of many hyperedges does not decrease sufﬁciently,making FM-based reﬁnement difﬁcult.Second,the weight of the vertices(i.e.,the number of vertices that have been collapsed together)in successively coarser graphs becomes signiﬁcantly different,which distorts the shape of the contracted hypergraph.To correct this problem,we implemented a MHEC scheme as follows.After the hyperedges to be contracted have been selected using the HEC scheme,the list of hyperedges is traversed again,and for each hyperedge that has not yet been contracted,the vertices that do not belong to any other contracted hyperedge are contracted together.B.Initial Partitioning PhaseDuring the initial partitioning phase,a bisection of the coarsest hypergraph is computed,such that it has a small cut, and satisﬁes a user-speciﬁed balance constraint.The balance constraint puts an upper bound on the difference between the relative size of the two partitions.Since this hypergraph has a very small number of vertices(usually less than200),the time toﬁnd a partitioning using any of the heuristic algorithms tends to be small.Note that it is not useful toﬁnd an optimal partition of this coarsest graph,as the initial partition will be sub-stantially modiﬁed during the reﬁnement phase.We used the following two algorithms for computing the initial partitioning. Theﬁrst algorithm simply creates a random bisection such that each part has roughly equal vertex weight.The second algorithm starts from a randomly selected vertex and grows a region around it in a breadth-ﬁrst fashion[22]until half of the vertices are in this region.The vertices belonging to the grown region are then assigned to theﬁrst part,and the rest of the vertices are assigned to the second part.After a partitioning is constructed using either of these algorithms,the partitioning is reﬁned using the FM reﬁnement algorithm.Since both algorithms are randomized,different runs give solutions of different quality.For this reason,we perform a small number of initial partitionings.At this point,we can select the best initial partitioning and project it to the original hypergraph,as described in Section II-C.However,the parti-tioning of the coarsest hypergraph that has the smallest cut may not necessarily be the one that will lead to the smallest cut in the original hypergraph.It is possible that another partitioning of the coarsest hypergraph(with a higher cut)will lead to a bet-KARYPIS et al.:MULTILEVEL HYPERGRAPH PARTITIONING:APPLICATIONS IN VLSI DOMAIN 73ter partitioning of the original hypergraph after the reﬁnement is performed during the uncoarsening phase.For this reason,instead of selecting a single initial partitioning (i.e.,the one with the smallest cut),we propagate all initial partitionings.Note that propagation of.Thus,by increasing the value ofis to drop unpromising partitionings as thehypergraph is uncoarsened.For example,one possibility is to propagate only those partitionings whose cuts arewithinissufﬁciently large,then all partitionings will be maintained and propagated in the entire reﬁnement phase.On the other hand,if the valueof,many partitionings may be available at the coarsest graph,but the number of such available partitionings will decrease as the graph is uncoarsened.This is useful for two reasons.First,it is more important to have many alternate partitionings at the coarser levels,as the size of the cut of a partitioning at a coarse level is a less accurate reﬂection of the size of the cut of the original ﬁnest level hypergraph.Second,reﬁnement is more expensive at the ﬁne levels,as these levels contain far more nodes than the coarse levels.Hence,by choosing an appropriate valueof(from 10%to a higher value such as 20%)did not signiﬁcantly improve the quality of the partitionings,although it did increase the run time.C.Uncoarsening and Reﬁnement PhaseDuring the uncoarsening phase,a partitioning of the coarser hypergraph is successively projected to the next-level ﬁner hypergraph,and a partitioning reﬁnement algorithm is used to reduce the cut set (and thus to improve the quality of the partitioning)without violating the user speciﬁed balance con-straints.Since the next-level ﬁner hypergraph has more degrees of freedom,such reﬁnement algorithms tend to improve the solution quality.We have implemented two different partitioning reﬁnement algorithms.The ﬁrst is the FM algorithm [8],which repeatedly moves vertices between partitions in order to improve the cut.The second algorithm,called hyperedge reﬁnement (HER),moves groups of vertices between partitions so that an entire hyperedge is removed from the cut.These algorithms are further described in the remainder of this section.1)FM:The partitioning reﬁnement algorithm by Fiduccia and Mattheyses [8]is iterative in nature.It starts with an initial partitioning of the hypergraph.In each iteration,it tries to ﬁnd subsets of vertices in each partition,such that moving them to other partitions improves the quality of the partitioning (i.e.,the number of hyperedges being cut decreases)and this does not violate the balance constraint.If such subsets exist,then the movement is performed and this becomes the partitioning for the next iteration.The algorithm continues by repeating the entire process.If it cannot ﬁnd such a subset,then the algorithm terminates since the partitioning is at a local minima and no further improvement can be made by this algorithm.In particular,for eachvertexto the other partition.Initially allvertices are unlocked ,i.e.,they are free to move to the other partition.The algorithm iteratively selects an unlockedvertex is moved,it is locked ,and the gain of the vertices adjacentto[8].For reﬁnement in the context of multilevel schemes,the initial partitioning obtained from the next level coarser graph is actually a very good partition.For this reason,we can make a number of optimizations to the original FM algorithm.The ﬁrst optimization limits the maximum number of passes performed by the FM algorithm to only two.This is because the greatest reduction in the cut is obtained during the ﬁrst or second pass and any subsequent passes only marginally improve the quality.Our experience has shown that this optimization signiﬁcantly improves the run time of FM without affecting the overall quality of the produced partitionings.The second optimization aborts each pass of the FM algorithm before actually moving all the vertices.The motivation behind this is that only a small fraction of the vertices being moved actually lead to a reduction in the cut and,after some point,the cut tends to increase as we move more vertices.When FM is applied to a random initial partitioning,it is quite likely that after a long sequence of bad moves,the algorithm will climb74IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI)SYSTEMS,VOL.7,NO.1,MARCH1999Fig.3.Effect of restricted coarsening .(a)Example hypergraph with a given partitioning with the required balance of 40/60.(b)Possible condensed version of (a).(c)Another condensed version of a hypergraph.out of a local minima and reach to a better cut.However,in the context of a multilevel scheme,a long sequence of cut-increasing moves rarely leads to a better local minima.For this reason,we stop each pass of the FM algorithm as soon as we haveperformedto be equal to 1%of the number ofvertices in the graph we are reﬁning.This modiﬁcation to FM,called early-exit FM (FM-EE),does not signiﬁcantly affect the quality of the ﬁnal partitioning,but it dramatically improves the run time (see Section IV).2)HER:One of the drawbacks of FM (and other similar vertex-based reﬁnement schemes)is that it is often unable to reﬁne hyperedges that have many nodes on both sides of the partitioning boundary.However,a reﬁnement scheme that moves all the vertices that belong to a hyperedge can potentially solve this problem.Our HER works as follows.It randomly visits all the hyperedges and,for each one that straddles the bisection,it determines if it can move a subset of the vertices incident on it,so that this hyperedge will become completely interior to a partition.In particular,consider ahyperedgebe the verticesofto partition 0.Now,depending on these gains and subject to balance constraints,it may move one of the twosets .In particular,if.III.M ULTIPHASE R EFINEMENT WITHR ESTRICTED C OARSENINGAlthough the multilevel paradigm is quite robust,random-ization is inherent in all three phases of the algorithm.In particular,the random choice of vertices to be matched in the coarsening phase can disallow certain hyperedge cuts,reducing reﬁnement in the uncoarsening phase.For example,consider the example hypergraph in Fig.3(a)and its two possible con-densed versions [Fig.3(b)and (c)]with the same partitioning.The version in Fig.3(b)is obtained by selectinghyperedgesto be compressed in the HEC phase and then selecting pairs ofnodesto be compressed inthe HEC phase and then selecting pairs ofnodesand apartitioningfor ,be the sequence of hypergraphsand partitionings.Given ahypergraphandorpartition,,are collapsedtogether to formvertexof,thenvertex belong。

热红外传感史

History of infrared detectorsA.ROGALSKI*Institute of Applied Physics, Military University of Technology, 2 Kaliskiego Str.,00–908 Warsaw, PolandThis paper overviews the history of infrared detector materials starting with Herschel’s experiment with thermometer on February11th,1800.Infrared detectors are in general used to detect,image,and measure patterns of the thermal heat radia−tion which all objects emit.At the beginning,their development was connected with thermal detectors,such as ther−mocouples and bolometers,which are still used today and which are generally sensitive to all infrared wavelengths and op−erate at room temperature.The second kind of detectors,called the photon detectors,was mainly developed during the20th Century to improve sensitivity and response time.These detectors have been extensively developed since the1940’s.Lead sulphide(PbS)was the first practical IR detector with sensitivity to infrared wavelengths up to~3μm.After World War II infrared detector technology development was and continues to be primarily driven by military applications.Discovery of variable band gap HgCdTe ternary alloy by Lawson and co−workers in1959opened a new area in IR detector technology and has provided an unprecedented degree of freedom in infrared detector design.Many of these advances were transferred to IR astronomy from Departments of Defence ter on civilian applications of infrared technology are frequently called“dual−use technology applications.”One should point out the growing utilisation of IR technologies in the civilian sphere based on the use of new materials and technologies,as well as the noticeable price decrease in these high cost tech−nologies.In the last four decades different types of detectors are combined with electronic readouts to make detector focal plane arrays(FPAs).Development in FPA technology has revolutionized infrared imaging.Progress in integrated circuit design and fabrication techniques has resulted in continued rapid growth in the size and performance of these solid state arrays.Keywords:thermal and photon detectors, lead salt detectors, HgCdTe detectors, microbolometers, focal plane arrays.Contents1.Introduction2.Historical perspective3.Classification of infrared detectors3.1.Photon detectors3.2.Thermal detectors4.Post−War activity5.HgCdTe era6.Alternative material systems6.1.InSb and InGaAs6.2.GaAs/AlGaAs quantum well superlattices6.3.InAs/GaInSb strained layer superlattices6.4.Hg−based alternatives to HgCdTe7.New revolution in thermal detectors8.Focal plane arrays – revolution in imaging systems8.1.Cooled FPAs8.2.Uncooled FPAs8.3.Readiness level of LWIR detector technologies9.SummaryReferences 1.IntroductionLooking back over the past1000years we notice that infra−red radiation(IR)itself was unknown until212years ago when Herschel’s experiment with thermometer and prism was first reported.Frederick William Herschel(1738–1822) was born in Hanover,Germany but emigrated to Britain at age19,where he became well known as both a musician and an astronomer.Herschel became most famous for the discovery of Uranus in1781(the first new planet found since antiquity)in addition to two of its major moons,Tita−nia and Oberon.He also discovered two moons of Saturn and infrared radiation.Herschel is also known for the twenty−four symphonies that he composed.W.Herschel made another milestone discovery–discov−ery of infrared light on February11th,1800.He studied the spectrum of sunlight with a prism[see Fig.1in Ref.1],mea−suring temperature of each colour.The detector consisted of liquid in a glass thermometer with a specially blackened bulb to absorb radiation.Herschel built a crude monochromator that used a thermometer as a detector,so that he could mea−sure the distribution of energy in sunlight and found that the highest temperature was just beyond the red,what we now call the infrared(‘below the red’,from the Latin‘infra’–be−OPTO−ELECTRONICS REVIEW20(3),279–308DOI: 10.2478/s11772−012−0037−7*e−mail: rogan@.pllow)–see Fig.1(b)[2].In April 1800he reported it to the Royal Society as dark heat (Ref.1,pp.288–290):Here the thermometer No.1rose 7degrees,in 10minu−tes,by an exposure to the full red coloured rays.I drew back the stand,till the centre of the ball of No.1was just at the vanishing of the red colour,so that half its ball was within,and half without,the visible rays of theAnd here the thermometerin 16minutes,degrees,when its centre was inch out of the raysof the sun.as had a rising of 9de−grees,and here the difference is almost too trifling to suppose,that latter situation of the thermometer was much beyond the maximum of the heating power;while,at the same time,the experiment sufficiently indi−cates,that the place inquired after need not be looked for at a greater distance.Making further experiments on what Herschel called the ‘calorific rays’that existed beyond the red part of the spec−trum,he found that they were reflected,refracted,absorbed and transmitted just like visible light [1,3,4].The early history of IR was reviewed about 50years ago in three well−known monographs [5–7].Many historical information can be also found in four papers published by Barr [3,4,8,9]and in more recently published monograph [10].Table 1summarises the historical development of infrared physics and technology [11,12].2.Historical perspectiveFor thirty years following Herschel’s discovery,very little progress was made beyond establishing that the infrared ra−diation obeyed the simplest laws of optics.Slow progress inthe study of infrared was caused by the lack of sensitive and accurate detectors –the experimenters were handicapped by the ordinary thermometer.However,towards the second de−cade of the 19th century,Thomas Johann Seebeck began to examine the junction behaviour of electrically conductive materials.In 1821he discovered that a small electric current will flow in a closed circuit of two dissimilar metallic con−ductors,when their junctions are kept at different tempera−tures [13].During that time,most physicists thought that ra−diant heat and light were different phenomena,and the dis−covery of Seebeck indirectly contributed to a revival of the debate on the nature of heat.Due to small output vol−tage of Seebeck’s junctions,some μV/K,the measurement of very small temperature differences were prevented.In 1829L.Nobili made the first thermocouple and improved electrical thermometer based on the thermoelectric effect discovered by Seebeck in 1826.Four years later,M.Melloni introduced the idea of connecting several bismuth−copper thermocouples in series,generating a higher and,therefore,measurable output voltage.It was at least 40times more sensitive than the best thermometer available and could de−tect the heat from a person at a distance of 30ft [8].The out−put voltage of such a thermopile structure linearly increases with the number of connected thermocouples.An example of thermopile’s prototype invented by Nobili is shown in Fig.2(a).It consists of twelve large bismuth and antimony elements.The elements were placed upright in a brass ring secured to an adjustable support,and were screened by a wooden disk with a 15−mm central aperture.Incomplete version of the Nobili−Melloni thermopile originally fitted with the brass cone−shaped tubes to collect ra−diant heat is shown in Fig.2(b).This instrument was much more sensi−tive than the thermometers previously used and became the most widely used detector of IR radiation for the next half century.The third member of the trio,Langley’s bolometer appea−red in 1880[7].Samuel Pierpont Langley (1834–1906)used two thin ribbons of platinum foil connected so as to form two arms of a Wheatstone bridge (see Fig.3)[15].This instrument enabled him to study solar irradiance far into its infrared region and to measure theintensityof solar radia−tion at various wavelengths [9,16,17].The bolometer’s sen−History of infrared detectorsFig.1.Herschel’s first experiment:A,B –the small stand,1,2,3–the thermometers upon it,C,D –the prism at the window,E –the spec−trum thrown upon the table,so as to bring the last quarter of an inch of the read colour upon the stand (after Ref.1).InsideSir FrederickWilliam Herschel (1738–1822)measures infrared light from the sun– artist’s impression (after Ref. 2).Fig.2.The Nobili−Meloni thermopiles:(a)thermopile’s prototype invented by Nobili (ca.1829),(b)incomplete version of the Nobili−−Melloni thermopile (ca.1831).Museo Galileo –Institute and Museum of the History of Science,Piazza dei Giudici 1,50122Florence, Italy (after Ref. 14).Table 1. Milestones in the development of infrared physics and technology (up−dated after Refs. 11 and 12)Year Event1800Discovery of the existence of thermal radiation in the invisible beyond the red by W. HERSCHEL1821Discovery of the thermoelectric effects using an antimony−copper pair by T.J. SEEBECK1830Thermal element for thermal radiation measurement by L. NOBILI1833Thermopile consisting of 10 in−line Sb−Bi thermal pairs by L. NOBILI and M. MELLONI1834Discovery of the PELTIER effect on a current−fed pair of two different conductors by J.C. PELTIER1835Formulation of the hypothesis that light and electromagnetic radiation are of the same nature by A.M. AMPERE1839Solar absorption spectrum of the atmosphere and the role of water vapour by M. MELLONI1840Discovery of the three atmospheric windows by J. HERSCHEL (son of W. HERSCHEL)1857Harmonization of the three thermoelectric effects (SEEBECK, PELTIER, THOMSON) by W. THOMSON (Lord KELVIN)1859Relationship between absorption and emission by G. KIRCHHOFF1864Theory of electromagnetic radiation by J.C. MAXWELL1873Discovery of photoconductive effect in selenium by W. SMITH1876Discovery of photovoltaic effect in selenium (photopiles) by W.G. ADAMS and A.E. DAY1879Empirical relationship between radiation intensity and temperature of a blackbody by J. STEFAN1880Study of absorption characteristics of the atmosphere through a Pt bolometer resistance by S.P. LANGLEY1883Study of transmission characteristics of IR−transparent materials by M. MELLONI1884Thermodynamic derivation of the STEFAN law by L. BOLTZMANN1887Observation of photoelectric effect in the ultraviolet by H. HERTZ1890J. ELSTER and H. GEITEL constructed a photoemissive detector consisted of an alkali−metal cathode1894, 1900Derivation of the wavelength relation of blackbody radiation by J.W. RAYEIGH and W. WIEN1900Discovery of quantum properties of light by M. PLANCK1903Temperature measurements of stars and planets using IR radiometry and spectrometry by W.W. COBLENTZ1905 A. EINSTEIN established the theory of photoelectricity1911R. ROSLING made the first television image tube on the principle of cathode ray tubes constructed by F. Braun in 18971914Application of bolometers for the remote exploration of people and aircrafts ( a man at 200 m and a plane at 1000 m)1917T.W. CASE developed the first infrared photoconductor from substance composed of thallium and sulphur1923W. SCHOTTKY established the theory of dry rectifiers1925V.K. ZWORYKIN made a television image tube (kinescope) then between 1925 and 1933, the first electronic camera with the aid of converter tube (iconoscope)1928Proposal of the idea of the electro−optical converter (including the multistage one) by G. HOLST, J.H. DE BOER, M.C. TEVES, and C.F. VEENEMANS1929L.R. KOHLER made a converter tube with a photocathode (Ag/O/Cs) sensitive in the near infrared1930IR direction finders based on PbS quantum detectors in the wavelength range 1.5–3.0 μm for military applications (GUDDEN, GÖRLICH and KUTSCHER), increased range in World War II to 30 km for ships and 7 km for tanks (3–5 μm)1934First IR image converter1939Development of the first IR display unit in the United States (Sniperscope, Snooperscope)1941R.S. OHL observed the photovoltaic effect shown by a p−n junction in a silicon1942G. EASTMAN (Kodak) offered the first film sensitive to the infrared1947Pneumatically acting, high−detectivity radiation detector by M.J.E. GOLAY1954First imaging cameras based on thermopiles (exposure time of 20 min per image) and on bolometers (4 min)1955Mass production start of IR seeker heads for IR guided rockets in the US (PbS and PbTe detectors, later InSb detectors for Sidewinder rockets)1957Discovery of HgCdTe ternary alloy as infrared detector material by W.D. LAWSON, S. NELSON, and A.S. YOUNG1961Discovery of extrinsic Ge:Hg and its application (linear array) in the first LWIR FLIR systems1965Mass production start of IR cameras for civil applications in Sweden (single−element sensors with optomechanical scanner: AGA Thermografiesystem 660)1970Discovery of charge−couple device (CCD) by W.S. BOYLE and G.E. SMITH1970Production start of IR sensor arrays (monolithic Si−arrays: R.A. SOREF 1968; IR−CCD: 1970; SCHOTTKY diode arrays: F.D.SHEPHERD and A.C. YANG 1973; IR−CMOS: 1980; SPRITE: T. ELIOTT 1981)1975Lunch of national programmes for making spatially high resolution observation systems in the infrared from multielement detectors integrated in a mini cooler (so−called first generation systems): common module (CM) in the United States, thermal imaging commonmodule (TICM) in Great Britain, syteme modulaire termique (SMT) in France1975First In bump hybrid infrared focal plane array1977Discovery of the broken−gap type−II InAs/GaSb superlattices by G.A. SAI−HALASZ, R. TSU, and L. ESAKI1980Development and production of second generation systems [cameras fitted with hybrid HgCdTe(InSb)/Si(readout) FPAs].First demonstration of two−colour back−to−back SWIR GaInAsP detector by J.C. CAMPBELL, A.G. DENTAI, T.P. LEE,and C.A. BURRUS1985Development and mass production of cameras fitted with Schottky diode FPAs (platinum silicide)1990Development and production of quantum well infrared photoconductor (QWIP) hybrid second generation systems1995Production start of IR cameras with uncooled FPAs (focal plane arrays; microbolometer−based and pyroelectric)2000Development and production of third generation infrared systemssitivity was much greater than that of contemporary thermo−piles which were little improved since their use by Melloni. Langley continued to develop his bolometer for the next20 years(400times more sensitive than his first efforts).His latest bolometer could detect the heat from a cow at a dis−tance of quarter of mile [9].From the above information results that at the beginning the development of the IR detectors was connected with ther−mal detectors.The first photon effect,photoconductive ef−fect,was discovered by Smith in1873when he experimented with selenium as an insulator for submarine cables[18].This discovery provided a fertile field of investigation for several decades,though most of the efforts were of doubtful quality. By1927,over1500articles and100patents were listed on photosensitive selenium[19].It should be mentioned that the literature of the early1900’s shows increasing interest in the application of infrared as solution to numerous problems[7].A special contribution of William Coblenz(1873–1962)to infrared radiometry and spectroscopy is marked by huge bib−liography containing hundreds of scientific publications, talks,and abstracts to his credit[20,21].In1915,W.Cob−lentz at the US National Bureau of Standards develops ther−mopile detectors,which he uses to measure the infrared radi−ation from110stars.However,the low sensitivity of early in−frared instruments prevented the detection of other near−IR sources.Work in infrared astronomy remained at a low level until breakthroughs in the development of new,sensitive infrared detectors were achieved in the late1950’s.The principle of photoemission was first demonstrated in1887when Hertz discovered that negatively charged par−ticles were emitted from a conductor if it was irradiated with ultraviolet[22].Further studies revealed that this effect could be produced with visible radiation using an alkali metal electrode [23].Rectifying properties of semiconductor−metal contact were discovered by Ferdinand Braun in1874[24],when he probed a naturally−occurring lead sulphide(galena)crystal with the point of a thin metal wire and noted that current flowed freely in one direction only.Next,Jagadis Chandra Bose demonstrated the use of galena−metal point contact to detect millimetre electromagnetic waves.In1901he filed a U.S patent for a point−contact semiconductor rectifier for detecting radio signals[25].This type of contact called cat’s whisker detector(sometimes also as crystal detector)played serious role in the initial phase of radio development.How−ever,this contact was not used in a radiation detector for the next several decades.Although crystal rectifiers allowed to fabricate simple radio sets,however,by the mid−1920s the predictable performance of vacuum−tubes replaced them in most radio applications.The period between World Wars I and II is marked by the development of photon detectors and image converters and by emergence of infrared spectroscopy as one of the key analytical techniques available to chemists.The image con−verter,developed on the eve of World War II,was of tre−mendous interest to the military because it enabled man to see in the dark.The first IR photoconductor was developed by Theodore W.Case in1917[26].He discovered that a substance com−posed of thallium and sulphur(Tl2S)exhibited photocon−ductivity.Supported by the US Army between1917and 1918,Case adapted these relatively unreliable detectors for use as sensors in an infrared signalling device[27].The pro−totype signalling system,consisting of a60−inch diameter searchlight as the source of radiation and a thallous sulphide detector at the focus of a24−inch diameter paraboloid mir−ror,sent messages18miles through what was described as ‘smoky atmosphere’in1917.However,instability of resis−tance in the presence of light or polarizing voltage,loss of responsivity due to over−exposure to light,high noise,slug−gish response and lack of reproducibility seemed to be inhe−rent weaknesses.Work was discontinued in1918;commu−nication by the detection of infrared radiation appeared dis−tinctly ter Case found that the addition of oxygen greatly enhanced the response [28].The idea of the electro−optical converter,including the multistage one,was proposed by Holst et al.in1928[29]. The first attempt to make the converter was not successful.A working tube consisted of a photocathode in close proxi−mity to a fluorescent screen was made by the authors in 1934 in Philips firm.In about1930,the appearance of the Cs−O−Ag photo−tube,with stable characteristics,to great extent discouraged further development of photoconductive cells until about 1940.The Cs−O−Ag photocathode(also called S−1)elabo−History of infrared detectorsFig.3.Longley’s bolometer(a)composed of two sets of thin plati−num strips(b),a Wheatstone bridge,a battery,and a galvanometer measuring electrical current (after Ref. 15 and 16).rated by Koller and Campbell[30]had a quantum efficiency two orders of magnitude above anything previously studied, and consequently a new era in photoemissive devices was inaugurated[31].In the same year,the Japanese scientists S. Asao and M.Suzuki reported a method for enhancing the sensitivity of silver in the S−1photocathode[32].Consisted of a layer of caesium on oxidized silver,S−1is sensitive with useful response in the near infrared,out to approxi−mately1.2μm,and the visible and ultraviolet region,down to0.3μm.Probably the most significant IR development in the United States during1930’s was the Radio Corporation of America(RCA)IR image tube.During World War II, near−IR(NIR)cathodes were coupled to visible phosphors to provide a NIR image converter.With the establishment of the National Defence Research Committee,the develop−ment of this tube was accelerated.In1942,the tube went into production as the RCA1P25image converter(see Fig.4).This was one of the tubes used during World War II as a part of the”Snooperscope”and”Sniperscope,”which were used for night observation with infrared sources of illumination.Since then various photocathodes have been developed including bialkali photocathodes for the visible region,multialkali photocathodes with high sensitivity ex−tending to the infrared region and alkali halide photocatho−des intended for ultraviolet detection.The early concepts of image intensification were not basically different from those today.However,the early devices suffered from two major deficiencies:poor photo−cathodes and poor ter development of both cathode and coupling technologies changed the image in−tensifier into much more useful device.The concept of image intensification by cascading stages was suggested independently by number of workers.In Great Britain,the work was directed toward proximity focused tubes,while in the United State and in Germany–to electrostatically focused tubes.A history of night vision imaging devices is given by Biberman and Sendall in monograph Electro−Opti−cal Imaging:System Performance and Modelling,SPIE Press,2000[10].The Biberman’s monograph describes the basic trends of infrared optoelectronics development in the USA,Great Britain,France,and Germany.Seven years later Ponomarenko and Filachev completed this monograph writ−ing the book Infrared Techniques and Electro−Optics in Russia:A History1946−2006,SPIE Press,about achieve−ments of IR techniques and electrooptics in the former USSR and Russia [33].In the early1930’s,interest in improved detectors began in Germany[27,34,35].In1933,Edgar W.Kutzscher at the University of Berlin,discovered that lead sulphide(from natural galena found in Sardinia)was photoconductive and had response to about3μm.B.Gudden at the University of Prague used evaporation techniques to develop sensitive PbS films.Work directed by Kutzscher,initially at the Uni−versity of Berlin and later at the Electroacustic Company in Kiel,dealt primarily with the chemical deposition approach to film formation.This work ultimately lead to the fabrica−tion of the most sensitive German detectors.These works were,of course,done under great secrecy and the results were not generally known until after1945.Lead sulphide photoconductors were brought to the manufacturing stage of development in Germany in about1943.Lead sulphide was the first practical infrared detector deployed in a variety of applications during the war.The most notable was the Kiel IV,an airborne IR system that had excellent range and which was produced at Carl Zeiss in Jena under the direction of Werner K. Weihe [6].In1941,Robert J.Cashman improved the technology of thallous sulphide detectors,which led to successful produc−tion[36,37].Cashman,after success with thallous sulphide detectors,concentrated his efforts on lead sulphide detec−tors,which were first produced in the United States at Northwestern University in1944.After World War II Cash−man found that other semiconductors of the lead salt family (PbSe and PbTe)showed promise as infrared detectors[38]. The early detector cells manufactured by Cashman are shown in Fig. 5.Fig.4.The original1P25image converter tube developed by the RCA(a).This device measures115×38mm overall and has7pins.It opera−tion is indicated by the schematic drawing (b).After1945,the wide−ranging German trajectory of research was essentially the direction continued in the USA, Great Britain and Soviet Union under military sponsorship after the war[27,39].Kutzscher’s facilities were captured by the Russians,thus providing the basis for early Soviet detector development.From1946,detector technology was rapidly disseminated to firms such as Mullard Ltd.in Southampton,UK,as part of war reparations,and some−times was accompanied by the valuable tacit knowledge of technical experts.E.W.Kutzscher,for example,was flown to Britain from Kiel after the war,and subsequently had an important influence on American developments when he joined Lockheed Aircraft Co.in Burbank,California as a research scientist.Although the fabrication methods developed for lead salt photoconductors was usually not completely under−stood,their properties are well established and reproducibi−lity could only be achieved after following well−tried reci−pes.Unlike most other semiconductor IR detectors,lead salt photoconductive materials are used in the form of polycrys−talline films approximately1μm thick and with individual crystallites ranging in size from approximately0.1–1.0μm. They are usually prepared by chemical deposition using empirical recipes,which generally yields better uniformity of response and more stable results than the evaporative methods.In order to obtain high−performance detectors, lead chalcogenide films need to be sensitized by oxidation. The oxidation may be carried out by using additives in the deposition bath,by post−deposition heat treatment in the presence of oxygen,or by chemical oxidation of the film. The effect of the oxidant is to introduce sensitizing centres and additional states into the bandgap and thereby increase the lifetime of the photoexcited holes in the p−type material.3.Classification of infrared detectorsObserving a history of the development of the IR detector technology after World War II,many materials have been investigated.A simple theorem,after Norton[40],can be stated:”All physical phenomena in the range of about0.1–1 eV will be proposed for IR detectors”.Among these effects are:thermoelectric power(thermocouples),change in elec−trical conductivity(bolometers),gas expansion(Golay cell), pyroelectricity(pyroelectric detectors),photon drag,Jose−phson effect(Josephson junctions,SQUIDs),internal emis−sion(PtSi Schottky barriers),fundamental absorption(in−trinsic photodetectors),impurity absorption(extrinsic pho−todetectors),low dimensional solids[superlattice(SL), quantum well(QW)and quantum dot(QD)detectors], different type of phase transitions, etc.Figure6gives approximate dates of significant develop−ment efforts for the materials mentioned.The years during World War II saw the origins of modern IR detector tech−nology.Recent success in applying infrared technology to remote sensing problems has been made possible by the successful development of high−performance infrared de−tectors over the last six decades.Photon IR technology com−bined with semiconductor material science,photolithogra−phy technology developed for integrated circuits,and the impetus of Cold War military preparedness have propelled extraordinary advances in IR capabilities within a short time period during the last century [41].The majority of optical detectors can be classified in two broad categories:photon detectors(also called quantum detectors) and thermal detectors.3.1.Photon detectorsIn photon detectors the radiation is absorbed within the material by interaction with electrons either bound to lattice atoms or to impurity atoms or with free electrons.The observed electrical output signal results from the changed electronic energy distribution.The photon detectors show a selective wavelength dependence of response per unit incident radiation power(see Fig.8).They exhibit both a good signal−to−noise performance and a very fast res−ponse.But to achieve this,the photon IR detectors require cryogenic cooling.This is necessary to prevent the thermalHistory of infrared detectorsFig.5.Cashman’s detector cells:(a)Tl2S cell(ca.1943):a grid of two intermeshing comb−line sets of conducting paths were first pro−vided and next the T2S was evaporated over the grid structure;(b) PbS cell(ca.1945)the PbS layer was evaporated on the wall of the tube on which electrical leads had been drawn with aquadag(afterRef. 38).。

1、下载文档前请自行甄别文档内容的完整性，平台不提供额外的编辑、内容补充、找答案等附加服务。
2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
3、如文档侵犯您的权益，请联系客服反馈,我们会尽快为您处理(人工客服工作时间：9:00-18:30)。

- Its all about Performance and Scalability
SQL REVIEW
1. Capture
2. Identify 3. Confirm 4. Analyze 5. Optimize 6. Verify
- Its all about Performance and Scalability
therefore
•Increase server throughput,
therefore
•Reduce H/W to run your
application
- Its all about Performance and Scalability
ABOUT Author
...
- Its all about Performance and Scalability
1. 2. 3. 4. 5. 6.
Capture Identify Confirm Analyze Optimize Verify
Capture
Example
Application Logging
24 statements on page Easy to spot worst query
175 ms 7 ms
TCP/IP Example
$ sudo tcpdump -i any port 3306 -s 65535 -x -nn -q -tttt -c 10000 | mk-query-digest --type tcpdump
function debug($str) { if (DEBUG) echo $str; return 0; } function format_time($time) { return number_format($time, 8, '.', ''); }
- Its all about Performance and Scalability
Capture SQL queries executed for 1 second Item
====== SELECT COMMIT SELECT SELECT SELECT
.. .. .. ..
/article/mk-query-digest
- Its all about Performance and Scalability
Six step process
Capture
General Query Log
TCP/IP Connectors Application Proxy
All Queries? or Sample Queries?
Not discussed in detail this presentation
Plugin
Duration
- Its all about Performance and Scalability
CONFIRM
Via
Application Logging mysql client SHOW PROFILES command
1. 2. 3. 4. 5. 6. Capture Identify Confirm Analyze Optimize Verify
QuickTime?and a GIF decompressor are needed to see this picture.
- Its all about Performance and Scalability
OBJECTIVE
•Improve query performance,
IDENTIFY
By Duration of time
By Frequency of execution
Not discussed in detail this presentation
- Its all about Performance and Scalability
Ronald BRADFORD
12 years with MySQL / 22 years with RDBMS
Senior Consultant at MySQL Inc (06-08)
Consultant for Oracle Corporation (96-99)
7 years presenting MySQL content
Application Logging Example
<?php PHP example mysql_query('SELECT col1, col2 FROM table1 ...'); ... ... mysql_query('SELECT a,b,c FROM x INNER JOIN y ... ... mysql_query('UPDATE table SET ... ?> // Create wrapper function function db_query($sql) { $rs = mysql_query($sql); return $rs;} // Global replace mysql_query with db_query
1. 2. 3. 4. 5. 6.
Capture Identify Confirm Analyze Optimize Verify
TCP/IP Example
Frequency
//TODO
# Profile # Rank Query ID Response time Calls R/Call Apdx V/M # ==== ================== ============= ===== ====== ==== ===== # 1 0xE5C8D4B9F7BDCCAF 0.5044 18.7% 1 0.5044 1.00 0.00 # 2 0x813031B8BBC3B329 0.4893 18.2% 23 0.0213 1.00 0.01 # 3 0x04AB3BC1A33D3B48 0.4107 15.2% 1 0.4107 1.00 0.00 # 4 0xD15CA257BAF77DAF 0.3356 12.5% 321 0.0010 1.00 0.00 # 5 0x228B1F36C5EBCF1F 0.2177 8.1% 2 0.1089 1.00 0.22 Item ====== SELECT COMMIT SELECT SELECT SELECT .. .. .. ..
All time top MySQor Oracle ACE Director
Available NOW for consulting

- Its all about Performance and Scalability
Application Logging Example
0.00010109 SELECT ip FROM ... 0.00005198 SELECT name, value FROM ... 0.00005984 SELECT id, status, ... 0.17592907 SELECT g.id, c.id, ... 0.00047803 SELECT DISTINCT id_c FROM camp... 0.00741315 SELECT DISTINCT id_c FROM camp.. 0.00058198 SELECT id_c FROM cr ... 0.00161815 SELECT id_c FROM cr ... 0.00032806 SELECT id_c FROM cr ... 0.00007200 SELECT DISTINCT id_a FROM arb ... 0.00005412 SELECT DISTINCT id_a FROM asw ... 0.00004697 SELECT id_adv FROM arw 0.00004601 SELECT id_adv FROM arw 0.00009012 SELECT gk.id, k.value ... 0.00009084 SELECT gk.id, k.value ... 0.00006318 SELECT gk.id, k.value ... 0.00005794 SELECT gk.id, k.value ... 0.00005603 SELECT gk.id, k.value ... 0.00006104 SELECT gk.id, k.value ... 0.00243688 SELECT gk.id, k.value ... - k.value Performance and Scalability SELECT gk.id, Its all about ...
- Its all about Performance and Scalability
Improving Performance with Better Indexes
Ronald Bradford OTN APAC Tour October 2011
- Its all about Performance and Scalability
Application Logging Example
function db_query($sql) { $start_time = microtime(true); $rs = mysql_query($sql); $exec_time = microtime(true) - $start_time; debug(format_time($exec_time) . ' ' . $sql . "\n"); return $rs; }
# Profile # Rank Query ID Response time Calls R/Call Apdx V/M # ==== ================== ============= ===== ====== ==== ===== # 1 0xE5C8D4B9F7BDCCAF 0.5044 18.7% 1 0.5044 1.00 0.00 # 2 0x813031B8BBC3B329 0.4893 18.2% 23 0.0213 1.00 0.01 # 3 0x04AB3BC1A33D3B48 0.4107 15.2% 1 0.4107 1.00 0.00 # 4 0xD15CA257BAF77DAF 0.3356 12.5% 321 0.0010 1.00 0.00 # 5 0x228B1F36C5EBCF1F 0.2177 8.1% 2 0.1089 1.00 0.22