Performance Prediction Of Parallel Applications On Networks Of Workstations

Networks of workstations NOWs are an alternative to supercomputers when it comes to running parallel applications In most environments users of a workstation use less than of the available CPU cycles The remaining capacity can be used to support tasks of a parallel application This paper presents a hybrid analytic and simulation model that can be used to predict the performance of parallel applications speci ed by a task graph on a network of workstations The model was validated through measurements and takes into account the speci c workstations used the scheduling discipline and the network characteristics Moreover it considers the e ect of the owner s interactive workload on the execution time of the parallel application The paper brie y discusses a tool developed by the authors to implement the model Introduction The idea of exploiting a Network of Workstations NOW to execute processes other than the ones submitted by the owner of the workstation arises from studies which have shown that a large percentage of a workstation CPU cycles are idle for over of the time on the average Al though the idea of a network of workstations itself is not new it is now possible because of recent technological ad vances in the areas of Networks new networking technologies like switched local area networks allow bandwidth to scale with the number of workstations and low overhead pro tocols have made fast communication possible Workstations present day workstations are very powerful machines A top end workstation today operates at almost one third the speed of a super computer or a massively parallel machine and costs less Also a typical workstation o ers large inex pensive memory and disk capacity which can be ac cessed across the network making the resources of a workstation useful and accessible to users other than the owner of the workstation There are several opportunities presented by network of workstations for di erent needs For NOWs to suc This work was supported in part by the Advanced Research Project Agency ARPA under contract DABT C ceed as an alternative to supercomputers and PCs they should deliver at least the interactive performance of a ded icated workstation while providing the aggregate resources of the network for sequential and parallel workloads which demand these resources This requires a resource alloca tion policy that explicitly preserves interactive performance while allowing dedicated and unused resources throughout the network to be used by demanding applications These resources include DRAMs for memory intensive programs disks for I O bound programs and CPU for parallel and distributed programs A careful resource allocation and scheduling policy is nec essary to ensure that idle cycles are easily located and al located to tasks that require them Although the principle behind global scheduling in a network of workstations is the same as in any distributed system the NOW environment poses additional constraints In a typical workstation en vironment jobs submitted by the owner of a workstation are typically interactive The owner submits a command to the workstation and waits for the reply After getting the reply the owner spends some time called the think time before issuing another command As mentioned ear lier the commands issued by the owner of a workstation use less than of the CPU time On the whole however the overall load distribution can vary greatly in magnitude over time and is not homogeneous in nature Hosts may be idle or running interactive applications or compute intensive batch jobs or both Also the set of ma chines which are available as hosts for remote execution is constantly changing These conditions make it di cult to employ static scheduling policies for job scheduling in a NOW As mentioned before the main requirement of a NOW to succeed as a Total Solution System is that it should deliver at least the interactive performance of a dedicated workstation while providing aggregate resources of the network for demanding applications Any scheduler for a Network of Workstations has to account for all the above factors and perform processor allocation accordingly Processor re allocation can be done by either preempt ing a running process or by not doing so In the case of preemptive transfers the process that is currently running is stopped transferred to another processor where it is restarted Condor and Stealth are two examples of schedulers for distributed applications in a network of workstations The performance of a parallel application running on a net work of workstations depends on the number and type of workstations used to support the application on the owner workload at each workstation on the network used to con nect the workstations and on the scheduling algorithm used to allocate parallel tasks to workstations This paper presents a hybrid analytic and simulation model that can be used to predict the performance of parallel applications speci ed by a task graph on a network of workstations The model takes into account the speci c workstations used the scheduling discipline and the network characteristics Moreover it takes into account the e ect of the owner s interactive workload on the execution time of the parallel application The paper brie y discusses a tool developed by the authors to implement the model The paper is organized as follows Section two presents some basic assumptions and de nitions Section three de scribes the performance model Section four presents the results of the experiments conducted to validate the model Section ve presents some performance prediction studies carried out with the validated model Section six presents a brief description of SimNOW a predictive performance an alyzer for NOWs designed and implemented by the authors Finally section seven presents some concluding remarks Basic Assumptions The environment considered here consists of W worksta tions connected through a local network Each workstation has an owner that generates an interactive workload Tasks belonging to a parallel application may be running at one or more workstations It is assumed that the processes gen erated by the owner of the workstation have preemptive priority over the parallel tasks running on it The interactive workload at workstation i is characterized as a closed workload composed of Ni processes that alternate between processing and think time periods The average duration of the think time at workstation i is Zi and the average CPU service demand of an owner generated pro cess at workstation i is Di I O is not considered in this study but could easily be included with straightforward ex tensions Our model allows for heterogeneous workstations A parallel programPi is characterized by the tuple T R D C where T fT Tng is a set of n tasks P f Ti Tj j Ti Tjg is a precedence relation that indicates which tasks depend on other tasks to be started D ft tng is the set of execution times of the n tasks It is assumed that these times are deterministic if it were not for the interference of the interactive workload C fci j j Ti Tj Pg and ci j is the communi cation demand between tasks Ti and Tj measured in number of bytes sent by Ti to Tj The precedence relation P is better depicted as a task graph where the nodes are the tasks of the parallel application and the arcs are elements of P Figure shows an example of a task graph The nodes of the graph have two labels the top one is the task number and the bottom one is the task execution time Arcs of the task graph are labeled with the communication demand As seen in the gure there are