Restructuring Distributed Algorithms for Mobile Computing Environment

Restructuring Distributed Algorithms for Mobile Computing EnvironmentArindam Chakraborty,98071 chak@cse.iitk.ac.in Ritesh Kumar,99324 riteshk@cse.iitk.ac.inManish Singh Badwal,98200msbadwal@cse.iitk.ac.induring CS634Mobile Computing course,under the supervision ofDr.R.K.Ghoshrkg@cse.iitk.ac.inIndian Institute of TechnologyDepartment of Computer Science and EngineeringKanpur208016,INDIAAbstractWe present a formalization of a subclass of distributed algorithms(which are used more often in a mobile computing environment)and give principles or thumbrules which must be followed so that the distributed algorithm executes well in themobile computing environment.In this report wefirst start with the classification of standard distributed algorithmsthat are most frequently used in distributed systems and look at how the two-tierprinciple discussed in B.R.Badrinath and Arup Acharya’s paper on restructuringdistributed algorithms,ports to these situations,and what modifications need tobe made.Then we discuss the pros and cons of completely generic restructuring(without requiring to modify the algorithm itself)to mobile-aware solutions wherethe algorithm has to explicitly handle the efects of mobility.We then stress on the importance of design abstraction of resources in a mobile en-vironment so that a greater amount of concurrency can be handled in the resource.To that effect we propose a method to evaluate a’concurrency load’for any ab-straction and use it to tell which of the abstractions would be better in the mobilesituation and in this relation we enumerate two principles which must be followed.1IntroductionDistributed algorithms are algorithms which generally invole more than one host in its execution.These generally are designed to work in networked environments though for describing techniques generally an abstraction(like a graph)is adopted.Someyears earlier,i.e.before mobile computing,distributed algorithms used to be designed keeping static hosts in mind.These algorithms generally execute by sending messages to each other over the network or whatever the communication medium is,which takes care of the synchronization and communication part between the hosts involved in the algorithm.The assumptions that are generally taken are that1.The hosts are static and there is very little cost in locating a host.The cost ismore of sending the message.2.Because the hosts are static,the cost of sending messages to the hosts isfixedand don’t change with time or other factors.rge bandwidth availability is assumed.4.It is also assumed that the hosts are not resource poor.Thefixed hosts haveenough computation power,memory etc.5.The hosts are also not short of power supply.Thus it is assumed that they remainon during the whole algorithm execution and their inability to recieve a message because of a power failure is often a cause of the failure of the algorithm.However,these assumptions must be relaxed in a mobile computing environment be-cause of the following reasons.1.The cost of locating a mobile host is often significant.There are extensive lo-cation management techniques which handle this.At any cost we would like to keepfinding location of a mobile host again and again to a minimum.The cost is taken to be equivalent,if not more,than sending a message(after lookup).2.The mobile environment also has persistence of connections irrespective of themobility of the host.This can cause change of the costs of sending messages to hosts during the algorithm is executing.3.The bandwidth available is very scanty.4.The mobile hosts are resource poor.Also,the amount of resources that can bepresent on the mobile host has a large variation.5.The mobile hosts work on battery and hence to conserve power they frequentlyhave to turn off their network interfaces.Such a mode of execution is called ’doze’mode or powersave mode.This means that we cannot afford to’wake’up the mobile host at unexpected times frequently.All this means that it poses a great challenge to somehow port the existing distributed algorithms to the mobile environment.However,here are a few observations1.As the mobile host is resource poor it is infeasible to do computation on the MH.Thus distributed algorithms which fragment the problem into smaller parts and execute it independently on many machines to get benefit in computing power are ruled out.It is simply because the mobile host is not designed to do a lot2of number crunching.Also,this may also mean requirement of a fast means of comuunication between mobile hosts themselves which is not present.Thus,we would like to disregard distributed algorithms of the above nature.2.Another important class of distributive algorithms deal with synchronization ofactivities.There would not be heavy number crunching code running on the hosts but the host would be participating in operations on globally available resources which require synchronization.This class of algorithms would be ideally be suit-able for restructuring in a mobile environment.This is because the mobile host user would primarily use his mobile host to access services than as mentioned in thefirst point run services on his mobile.One of such projects is the“Wireless Coyote”[4].It is not that the distributed algorithms would not run in the mobile environment.It is only that they would be highly inefficient and costly due to the nature of the mobile computing environment.In fact work on restructuring distributed algorithms generally tend to keep the mobile host away from any computation.The distributed algorithm actually runs on the static hosts only with only the initiation of the algorithm and con-veying results being only place where moile hosts come in.Other than this the only difference is of synchronization between the mobile hosts which is handled by the fixed host acting as a proxy for the mobile host.This principle is well explained in B.R.Badrinath and Arup Acharya’s papers on restructuring distributed algorithms for mobile computing environments[1][2][3].We would like to give another formulation for distributed algorithms in a mobile computing envirnment.However,we would restrict ourselves those classes of algo-rithms which donot run distributed services but those which deal with synchronization issues in accessing a service from many hosts.A typical example of such a service may be a distributedfilesystem.2Issues involved in porting distributed algorithms to mobile environmentsLet us look at the issue we are faced with when restructuring our class of algorithms into a mobile computing environment.We basically have to take care of two things.Bandwidth and resourcesSynchronization2.1Bandwidth and resourcesWhenever we have an operation to be executed on a remote object(the obvious case in a mobile computing environment)we would have to send messages to the host which hosts the object.Our operations would be performed then by the remote host on our behalf(Though it is some times convenient to think of the host’logically’executing the operation directly on the remote host).Whenever we require to send messages we would require two things31.Figure out whom to send the message.This operation is exceptionally costly ina mobile computing environment and is handled by extensive location manage-ment techniques.2.Actually send the message.If the message was supposed to be sent to afixed host the above two activities are carried out by the MSS of the mobile host in thefixed network.This is a lot less costly than the case when the message is sent to another mobile host.This is because no location search is required in this case(unless the MH explicitly moves from the MSS’s cell in which case the MSS would have to search for the MH to send it the message back).Thus its best if we restrict sending messages to mobile hosts except that case when the MSS has to send one to one of its MHs.However this has important implication on the objects and operations model we are developing.As mobile hosts are not advised to send messages to another mobile hosts we indirectly are saying that a mobile host should not execute operations on an object on another mobile host.Basically it gives us a principle:”We should not have objects (which are remotely accessed)on mobile hosts.They must be present only onfixed hosts.”We may note that this is also in accordance with the two tier principle illustrated in B.R.Badrinath and Arup Acharya’s structuring of distributed algorithms[3].The very fact that a mobile host is poor in computing and power resources,has low band-width and high cost location search caused the two tier principle to say that as much as possible all computation work should be done onfixed hosts leaving the least for the mobile host.In our case an object is associated with something which requires com-puting power and bandwidth to exist and we are disallowing such things on the mobile host.The role of the mobile host comes is being involved in the thread of execution which executes operations on the object.2.2SynchronizationWhenever we have a centrally located resource which is accessed from a number of places there is contention.Let us call our centrally located resource as an object.Each sequence of operations that are being initiated from a specific place(a mobile host) can be called a’thread’of execution.Thus we many concurrently running threads of execution each of which are trying to run operations on an object.For the time being let us take no assumptions of where this object can be.It can be on afixed host or it may also reside on a mobile host.Whenever we have concurrent threads operating on the same object there is contention.The threads compete for gaining access to the object.As we donot want the threads to run operations on objects in an inconsistent state we would like to synchronize the operations on the object.We would later discuss what would be a formal way to model the synchronization primitives.43Bandwidth and ResourcesWe start by stating the two-tier principle and propounded by Badrinath and Acharya in their paper on restructuring distributed algorithms.To the extent possible,computation and communication costs of an algorithm is borne by the static portion of the network.This attempts to avoid locating a mobile par-ticipant and lowers the”search cost”of the algorithm;additionally,the number of operations performed at the mobile hosts and thereby,consumption of battery power, which is a critical resource for mobile hosts,is kept to a minimum.The general pattern of all distributed algorithms can be visualised as comprising a com-munication component and calculation component.Being a distributed execution,the calculation component is limited to the individual hosts and during these calculations it might require to communicate with its neighbours or other nodes,so as to exchange information regarding the state of its or other’s registers,the results of current compu-tations,contacting some host for ordering or maintenance tasks in the algorithm or to get initialization parameters for the next phase of computation and so on.Thus it is nat-ural to expect that a distributed system of hosts would show bursts of communication in between periods of local computation.This is the model wefind in the context of fixed hosts but once the possibility of mobile hosts comes into the picture,we are faced with the problems of”communication costs”,”location detection of mobile host”and ”scarcity of computational resources at mobile hosts”.To deal with the later problem, we have to try and schedule all the computation intensive parts of the algorithm to the fixed hosts,to as great an extent as possible.To address thefirst two issues,we will try to adapt the”search”,”inform”and”proxy”strategies introduced in Badrinath and Acharya’s paper in relation to restructuring a logical token ring network.These princi-ples are in fact more generally and can be used in various other classes of distributed algorithms,as we shall shortly describe.First we briefly look at what these strategies are:1.Search The Search strategy basically entails that each MSS(Mobile Service Sta-tion)is responsible for the MHs(Mobile Hosts)in its cell.So all the communi-cation primitives are executed by the MSS on behalf of the MH and the results of the communication are then communicated to the MH.However if an MH changes its location so that it moves to a different MSS jurisdiction then a search has to be done in the entire network to locate the current location of the MH.This requires broadcasting the location query for the MH to all thefixed nodes (MSSs)and then waiting for a reply from the MSS under which the MH is now located.rm The Inform strategy is similar to the Search,with difference that the mo-bile host is expected to inform the MSS of its movement from one cell to another.Thus when the MH moves from MSS1to MSS2,it informs MSS2of the identity of its previous MSS(in this case MSS1)and then MSS2informs MSS1of the new location for the MH.53.Proxy Proxy combines both the above strategies in the sense that the set of allMSSs are partitioned into various regions or sets and a static host proxy made responsible for the region.In the case of a wide area move(the proxy under which the mobile host is,changes),the MH informs the proxy of its location change,while in the case of a local area move(proxy doesnot change),normal search is used.The Search strategy works better if the mobile host shows high mobility,so that the uncertainty in its position is so high,that it is easier to search for it.On the other hand for a less mobile host Inform will perform better because extensive search would be too expensive.We now look at the different categories of distributed algorithms with example of each,and see what generic restructuring can be done for these categories.Non-coordinator based systems–Systems with all machines equivalent–Systems with exception machinesCoordinator based systems–Fixed coordinator systems–Moving coordinator systemsLet us look into each of these categories one by one,along with suitable examples.3.1Non-coordinator systems with all machines equivalentIn these systems all machines are absolutely equivalent in the sense that the amount of computational and communicational load shared by each machine is same.Such systems can pretty easily be modified to work excellently in a mobile computing en-vironment.We illustrate the method with Lamport’s Bakery algorithm for mutual exclusion as an example.In the Bakery algorithm,a process waiting to enter its critical section chooses a number.This number must be greater than all other numbers currently in use.There is a global shared array of current numbers for each process.The entering process checks all other processes sequentially,and waits for each one which has a lower number.Ties are possible;these are resolved using process IDs.However to look at the impact of executing this directly on the mobile hosts we donot have to look into the details of the algorithm execution,but only at the communication aspects of the algorithm.Note that since the computation load of each machine is equivalent hence there isn’t much we can improve in that respect.So in the algorithm,each machine has to access the registers of the other machines.The drawbacks of non-restructuring include1.High search cost:Since every message exchange is between mobile hosts henceevery message is likely to incur a search cost.This leads to a high search costwhich is proportional to the number of mobile hosts in the system(N).62.Battery consumption at mobile hosts will also be high.In each step the overallenergy consumption is again going to be proportional to the number of mobile hosts.3.Since the algorithm requires the participation of each mobile host in each host,hence it doesnot allow any host to disconnect or doze in an interrupted manner.4.Correctness of the algorithm requires that messages be delivered in Fifo orderand this places a burden on the underlying network to maintain a logical Fifo channel between any pair of mobile hosts.In order to improve the overheads of the bakery algorithm,we will have to shift the communication burden onto thefixed hosts.So instead of N mobile hosts executing the algorithm,it is the MSSs that will maintain the necessary data structures for the al-gorithm(request queue,request and reply messages)and thus secure mutual exclusion on behalf of the mobile hosts.The necessary modifications thus will be as follows:1.Time stamping is used only for messages exchanged between the MSSs(whichnow form the components of bakery algorithm)and not for messages exchanged between the mobile hosts.2.A mobile host initiates the algorithm by sending an initialization message to itslocal MSSwhich will now execute the algorithm on behalf of the mobile host.Thus the appropriate request,reply and release messages are marked with the id of the mobile host for which it is sent by the MSS.3.When the MSS secures mutual exclusion for the MH(using normal bakery al-gorithm execution),it informs the mobile host of the same.But since the mobile host might have changed its location,hence potentially a search will be required.4.The mobile host then accesses the critical section and informs the MSS that itsaccess is complete and hence a release message is sent to every other MSS.5.If the mobile host disconnects prior to the requested access to the critical sec-tion being received,then every disconnection has to be informed to the MSS,so that on receiving the grant for access,the MSS will instantly make the release response,so that other requests are not delayed.Since timestamping of messages is used and each MSS maintains a queue of pend-ing requests,hencefifo ordering of messages is easy to ensure.Also as the MSSs execute the algorithm without any modifications and messages for different MHs are distinguished by tagging the host ids to them,thus correctness of the algorithm is en-sured.Overall communication cost of this algorithm will be as follows:Initialization requires a,granting a request requires,and release of lock requires.The overall cost(alongwith the cost for executing the algorithm between MSSs)thus becomeswhere M is the number of MSSs(partcipating in the algorithm execution).7Note that comparing with the previous approach where the bakery algorithm was executed directly at the mobile hosts,since and,thus the overall communication cost is much less in the later case.Mutual exclusion is also achieved using a logical ring structure(hosts access the critical section only when they receive the token in its due course of traversal in the ring).This algorithm translates to distributed case in exactly the same manner as above, alongwith the same associated benefits in terms of communication cost savings.3.2Non-coordinated systems with exception machinesSuch systems are similar to the previous category,and differ only in the sense that the code executed by one machine(or a few of them)is different than that executed by most of the other machines,and thus these machines constitute the”exception”machines in that sense.An example of this category is dijkstra’s self stabilizing algorithm.Consider a system consisting of a set of nfinite state machines connected in the form of a ring,with a circulating privilege which denotes the ability of a machine to change its state.when a machine has the privilege,it is able to change its current state which is termed a move.Thus the system is self-stabilizing when,regardless of the initial state and privilege selected each time,the system always converges to a legal configuration in afinite number of steps.On the presence of multiple privileges in the system,the choice of whihc machine is entitled to make the move can be decided arbitrarily.A legal state of the system has the following properties:There must be atleast one privilege in the system(no deadlock).Every move from a legal state must again put the system into a legal state(clo-sure).During an infinite execution,each machine should enjoy the privilege an infinite number of times(no starvation).Given any two legal states,there is a series of moves that change one legal state to another(reachability).Let us now look at Dijkstra’s algorithm involving K states where where n is the total number machines.For any machine,we use the symbols S,L,R to denote its own state,the state of left neighbour,and the state of the right neighbour respectively.the exception machineif then.the other machinesif then.Similarly a better algorithm(also by Dijkstra)using3states and two exception ma-chines is as follows:the bottom machine,machine0if then.8the top machine,machine n-1if and then.the other machinesif thenif then.The details of how the self-stabilization works is however not very important so far as porting the algorithms to mobile environment is concerned.Instead we have to look at the communication that occurs between two machines.In both the algorithms cited and also other self-stabilization algorithms,the essen-tial communication perspective is to access the registers for the left and right neigh-bours.Thus clealry these class of algorithms are very similar to the previous class of algorithms and the presence of one or more exception machines really doesnot make much of a difference to the communication costs involved.Thus depending on the mo-bility of the hosts,either of SEARCH,INFORM or PROXY strategies can be used to adapt to the mobile environment.In fact since in this case only the neighbour’s values are needed,hence the overhead of wireless communication will be even less,as most’neighbouring’hosts will be under the same MSS except for two(at the edges).3.3Fixed coordinator based systemDistributed algorithms which involve a coordinator for running the algorithm properly induce a greater than normal communication overhead on the coordinator.Thefirst class of such algorithms is thefixed coordinator set of algorithms which involves one particular host being permanently assigned to be the coordinator in the set of machines or processes.Thus apart from the normal optmization possible for the mobile hosts,the communication pattern for the coordinator has to be specifically optimized,which will yield better dividends in terms of reducing the communication cost,because a large portion of the communication in a coordinator based system will be centered around the coordinator.An example of thefixed coordinator system is encountered in the case of total order atomic broadcast algorithms.The system here consists of N hosts,each of which par-ticipates in the process by broadcasting messages to the other hosts.After a broadcast is made,the hosts on receiving the message sends the messages with their timestamps to the coordinator thread.On receiving the relayed broadcasted message from all the processes,the coordinator sets the timestamp of the message to the maxima of the re-ceived timestamps.This is then broadcasted back to the hosts.The total ordering on the broadcasted messages is induced by this timestamp given by the coordinator(figure 1).So now in the case offixed coordinator algorithms,the coordinator is alwaysfixed to a particular host(in thefigure host3).In modifying such algorithms for execution in a mobile environment,we have to give special emphasis to the coordinator.The following modifications will be required:1.If possible,thefixed coordinator is executed on a static host,that is an MSS.9timehost 1host 3host 4host 5host 2Figure 1:Total Order Atomic Broadcast using fixed/moving coordinator.2.The other hosts will be mobile and the normal search,inform and proxy strategies can be applied depending on the mobility characterestics of the system.In the case where the algorithm is directly executed on the mobile hosts without any change,the total cost incurred for each broadcast will be,cost of broadcast (),cost of sending to the coordinator (),broadcast back of timestamped message ()leading to a total cost ofThis can be used slightly improved if the search results are cached (particularly for the coordinator)by each MSS:However on following the above strategy,the cost becomes cost of broadcast (),cost of sending to coordinator (),cost of broad-casting back the final timestamped message ()making it a total cost ofThus restructuring in the normal manner alongwith placing the coordinator on an MSS leads to great savings.3.4Moving coordinator based systemThe difference of a moving coordinator system from the afore-mentioned fixed se-quencer one is in the sense that the coordinator can now move between the various mobile hosts in the system.Thus who is the coordinator is a function of time as shown in figure 2.Thus in this case normal algorithm execution at mobile hosts will again take the same amount of complexity as in the previous analysis.However we can modify the system as follows:10Figure2:Conceptual model of a moving coordinator system,the sender,coordina-tor and receiver sets are the same,that is comprise of the same hosts,but are shown distinctly for the sake of clarity.1.The normal search,inform or proxy strategies are used for the all the nodes inthe system.2.As soon as a monile host becomes a coordinator,then the number of accesses toit will rise drastically in a short space of time,hence the MH informs its MSS of its change of status to coordinator,which is then broadcasted to all MSSs.Also the MH during the duration of its tenure as coordinator usses the inform strategy, while other hosts use the search strategy.Using these modifications,each step of the algorithm will now require cost of broad-cast(),cost of sending to coordinator(),cost of broadcasting back thefinal timestamped message()and the additional overhead associated with change of coordinator().Thus the total cost comes towhere a change of coordinator occurs every paring this with the normal unrestructured costwe see that the savings are significant.4Concluding remarks for bandwidth and resource op-timizationA distributed algorithm for static hosts can be ex-tended to cover MHs in a uniform manner as follows:a proxy is associated with a MH for the duration of the MH’s11lifetime;a proxy will be informed about the location of its MH(s)whenever such a MH changes its location,and will be responsible for receiving and sending messages to the MHs associated with it.Now,the distributed algorithm can be extended to the mobile environment by executing the algorithms at the proxies of the participating mobile hosts.This provides a two-layer structure in which one layer executes the algorithm over the set of static hosts(proxies)and the other layer handles host mobility, i.e.the interaction between a proxy and the MHs”under”it.In this case,with a fixed association between a MH and its proxy,a total separation of mobility from the algorithm is achieved.However,this is not always a desirable solution because a proxy has to be informed of every move by a MH;in case of”wide area moves”and for MHs that frequently change their cell,this leads to a high message traffic from the MH to its proxy and may be practically infeasible.Thus,we need to look for less static solutions in which the association between the MHs and proxies change,depending on the mobility of hosts.One way offlexibly associating a proxy with a MH is to explicitly incorporate the mobility of hosts into the structure of the algorithm itself.The search and inform strategies reflect this approach.The association between a MH and its proxy changes on everymove:the local MSS acts as a proxy to the MHs in its cell.Thus,a MH need not inform any non-local MSS of its current location.However,the obligations of a proxy may involve searching for a MH,and needs to be carefully defined for correctness of the algorithm.An intermediate approach between a lifelong asso-ciation of a proxy with a MH and assigning a new proxy on every move,is to consider a logical hierarchy offixedhosts, called”location servers”,such that the MHs are the leaves that attach to this hierarchy: MSSs are at level1in the hierarchy.The proxy for a MH is now defined to be a location-server that is d levels higher than a leaf.In this case,the association between a MH and its proxy remains invariant as long as the MH movesbetween locations that have a common d-level ancestor.A distributed algorithm designed for this setup will need to explicitly handle the effects of mobility,since the association between a MH and its proxy could change;however,compared to the approach where the association between a MH and its proxy is redefined on every move,the association does not change on every move in this approach.When the value of d is1,then this approach is equivalent to associating the local MSS as the proxy of a MH,with the association being redefined on every move;when d equals the number of levels in the hierarchy, then there is only one proxy for all MHs,that is a centralized and static solution.Note that a d=2reflects the Proxy strategy suggested by Badrinath and Acharya.5SynchronizationBefore we proceed let us restate our abstractions again.We have objects which are resources which are remotely accessed.We would not worry about’remotely’here but rather say that concurrent threads of execution execute operations on objects.The mobile host may initiate one or more such threads of execution.12。