Nguyen TD, Vaswani R, Zahorjan J. Using runtime measured workload characteristics in parallel processor scheduling. Science, vol. 1162). Springer: Berlin, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... and are executed in a simple First Come First Served policy (FCFS) and that they are malleable [Feitelson97] We propose to use dynamically measured application efficiency to improve the performance of two scheduling policies proposed so far, the equipartition [McCan93] and the equal efficiency [NguyenZV96]. The main idea is to allocate more processors to those parallel applications that achieve a target efficiency, allocating processors proportionally to the application s performance. The goal of this work is to show that by introducing small modifications in a scheduling policy we can consider the ....

....either there are no remaining jobs or until all P processors have been allocated. The only information provided by the application is the maximum number of processors that it can use. Reallocations are done at job arrival and completion. 3. 2 Equal efficiency The goal of the equal efficiency [NguyenZV96] is to maximize the system efficiency. The idea is to allocate more processors to those applications that have better efficiency and fewer processors to applications with worse efficiency. The equal efficiency initially assumes that all the applications have the same efficiency, then it allocates ....

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". JSPP, vol. 1162 of Lectures Notes in Computer Science. Springer-Verlag, Univ. of Washington, 1996.

....of the application, decisions can be taken dynamically leading to dynamic optimization. Dynamic performance measurement of applications is useful for several purposes. Based on the run time performance of applications the resources of the system may be allocated differently. In [Corbalan2000] [NguyenZV96] authors propose to consider the run time calculated application efficiency to perform the processor allocation. Moreover, performance measurements can be useful to dynamically optimize the application execution. In [Voss99] VossEigenmann99] authors propose to perform several run time tests to ....

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". JSPP, vol. 1162 of Lectures Notes in Computer Science. Springer-Verlag, University of Washington, 1996.

....basis, and use that information. All identifying characteristics associated with the submital of a job can potentially be used to determine its class. These characteristics include the user id, the file to be executed, the memory size specified, and possibly others. An estimate of the efficiency [59] or the execution time [31] of a job being scheduled can be obtained from retained statistics on the actual resource usage of jobs from the same (or a similar) class that have been previously submitted and executed. A small database can be kept to record resource consumption of jobs on a class by ....

....with. Second, even with very many queues, the resolution in which requirements are expressed is necessarily very coarse, and user estimates are notoriously inaccurate anyway. Recent more theoretical work shows how data can be acquired automatically by the system, rather than relying on the users [59,31,12]. At the same time that theoretical work is focusing, at least to some degree, on practical concerns, practice in the field seems to be rather oblivious of this development. One reason is that the larger and more advanced installations have been developing rather elaborate scheduling facilities, ....

T. D. Nguyen, R. Vaswani, and J. Zahorjan, "Using runtime measured workload characteristics in parallel processor scheduling ". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 155--174, SpringerVerlag, 1996. Lecture Notes in Computer Science Vol. 1162.

....particular memory mapping, etc. In [Corbalan99] we propose a new approach to dynamically calculate the speedup of parallel applications based on the relationship between an execution time baseline (measured at run time) and the execution time with the number of allocated processors. Nguyen et al. [NguyenZV96] propose an approach to dynamically measure the 2 of 14 efficiency of parallel applications based on the sources of overhead: idleness, communication (remote memory accesses) and system overhead. Smith et al. in [Smith99] propose a run time predictor to estimate queue wait times and improve ....

.... submitted to a queueing system and are executed in a simple First Come First Served policy (FCFS) In this work we propose to use dynamically measured application efficiency to improve the performance of two scheduling policies proposed so far, the equipartition [McCan93] and the equal efficiency [NguyenZV96]. The main idea is to allocate processors only to those parallel applications that achieve a target efficiency, allocating processors proportionally to the application s performance. The goal of this work is to show that by introducing small modifications in a scheduling policy we can consider ....

[Article contains additional citation context not shown here]

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". Job Scheduling Strategies for Parallel Processing, vol. 1162 of Lectures Notes in Computer Science. Springer-Verlag, University of Washington, 1996.

....amount of work of the system. These parameters are either already known or can be directly measured by the scheduler. 2 of 15 However, many researchers have shown that using application characteristics such as the speedup or the average parallelism improve the performance of the scheduler [2] 3][12]. In particular, Parsons and Sevcik [13] showed that, if applications executing in a system have different speedup curves, the knowledge of this behavior is useful to the scheduler, since it can assign more processors to those applications that will take advantage of them. The speedup is the ....

....the methodology that allows a dynamic computation of the speedup, and finally, we discuss some implementation issues. 3.1 Applications A great number of scientific applications are characterized by their predictable behavior. These applications are known as iterative parallel applications [12]. Figure 2 shows the structure of these applications. We can observe that they are composed of a set of parallel loops inside a sequential loop. We refer to this structure as iterative parallel region, and, to the set of parallel loops inside the sequential loop as parallel region. Since a ....

[Article contains additional citation context not shown here]

T. D. Nguyen,J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". Job Scheduling Strategies for Parallel Processing, volume 1162 of Lectures Notes in Computer Science. Springer-Verlag 96. University of Washington

....assume that jobs are capable of adapting to changes in the number of processors that are allocated to them. Adaptive programming techniques and runtime support for program restructuring are active areas of research and appear to be feasible for both shared memory and message passing systems (e.g. [27, 20, 6]) Although job reconfiguration can involve substantial cost, particularly if massive data movement is required, the results in [6, 11] show that the benefit of better processor scheduling can outweigh the associated cost. This key issue is explored further for EQS and FB PWS in section 4. We ....

.... number of processing nodes that each job can make productive use of, either because this information is specified when the job is submitted or because the system is capable of determining this information at runtime using methods such as the self tuning approach recently proposed by Nguyen et al. [20]. Similarly, for the FB PWS policy (or the EQS PWS policy yet to be defined) we assume that the system is capable of knowing the pws measure for each job, perhaps from runtime estimation techniques similar to those described in [20] In the remainder of the paper we assume the set of processing ....

[Article contains additional citation context not shown here]

T. D. Nguyen, R. Vaswani, J. Zahorjan, Using Runtime Measured Workload Characteristics in Parallel Processor Scheduling. Proc. IPPS '96 Workshop on Job Scheduling Strategies for Parallel Systems, Honolulu, Hawaii, April 1996.

....be useful and suitable for the incorporation in dynamic optimization tools. 1 Introduction Dynamic performance measurement of applications is useful for several purposes. Based on the run time performance of applications the resources of the system may be allocated differently. In [Corbalan2000][NguyenZV96] authors propose to consider the run time calculated application efficiency to perform the processor allocation. Moreover, performance measurements can be useful to dynamically optimize the application execution. In [Voss99] VossEigenmann99] authors propose to perform several run time tests to ....

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". Job Scheduling Strategies for Parallel Processing, vol. 1162 of Lectures Notes in Computer Science. Springer-Verlag, University of Washington, 1996.

.... can outweigh the benefits of the processor reallocation itself, as has been demonstrated both analytically and experimentally [6, 17] Sophisticated run time systems are required to do dynamic data redistribution automatically, and keep it from being an additional burden on the application writer [3, 8, 16]. Recently, it has been shown by Feitelson and Nitzberg [9] that application characteristics, such as execution time on a given number of processors, can be estimated. These estimates can be exploited to achieve improved scheduling. In their measurement studies, the majority of jobs run repeatedly ....

....function for a dynamic policy is not as straightforward as in the case of static partitioning, due to the continuous change in the number of processors. But both SDF and DIF need some form of execution time representation in order to be realized. The approximation method proposed by Nguyen et al. [16] for shared memory machines, where 24 the best partition size of an application is determined by sample executions on different numbers of processors, induces overhead which can become non negligible in distributed memory systems. Also, by comparing our algorithm AEP(2) with ASP(2) we realize ....

Thu D. Nguyen, Raj Vaswani, and John Zahorjan. Using Runtime Measured Workload Characteristics in Parallel Processor Scheduling. Technical Report UW-CSE-95-10-01, Department of Computer Science and Engineering, University of Washington, Seattle, October 1995.

....allocation may be changed after it has begun execution, a feature that normally requires explicit support within the application. 2 In the process control approach, the application must be designed to to adapt dynamically to changes in processor allocation while it is running [TG89,GTS91,NVZ96] As this type of support is uncommon, a simpler strategy may be to rely on application level checkpointing, often used by long running jobs to tolerate system failures. For these cases, it might be possible to modify the application so as to store checkpoints in a format that is independent of ....

Thu D. Nguyen, Raj Vaswani, and John Zahorjan. Using runtime measured workload characteristics in parallel processor scheduling. In Dror G. Feitelson and Larry Rudolph, editors, Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science Vol. 1162, pages 175--199. Springer-Verlag, 1996.

....Origin2000 [29] This work attacks the problem of the processor allocation in an execution environment where no knowledge of the application is available when applications are submitted. Many researchers have considered the use of application characteristics in processor scheduling [3] 6] 17] 22][23][24] In these works, parallel applications are characterized by different parameters such as the maximum speedup, the average parallelism, or the size of the working set. Performing the processor allocation without taking into account these characteristics can result in a bad utilization of the ....

....applications, and assuming a priori information. Some researchers propose that applications should monitor themselves and tune their parallelism, based on their performance. Voss et al. [31] propose to dynamically detect parallel loops dominated by overheads and to serialize them. Nguyen et al. [22][23] propose SelfTuning, to dynamically measure the efficiency achieved in iterative parallel regions and select the best number of processors to execute them considering the efficiency. These works have demonstrated the usefulness to use run time information. Other authors propose to communicate ....

[Article contains additional citation context not shown here]

T. D. Nguyen,J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". Job Scheduling Strategies for Parallel Processing, volume 1162 of Lectures Notes in Computer Science. Springer-Verlag 96. University of Washington

.... have been extensive empirical studies on the preemptive cost caused by time space sharing scheduling policies [38] 20] 37] Even for some cases when the preemption cost is relatively high, simulation and experimental studies support preemptive over non preemptive scheduling policies [38] 20] [23] [24] 1.2. Competitive Analysis. We make the assumption that job execution times are not known prior to their completion. This is quite realistic for modern general purpose multiprocessors. Since execution times are not known at the time jobs are scheduled, it is possible that any given ....

.... non preemptive algorithms has been subsequently Preemptive Scheduling of Parallel Jobs on Multiprocessors 5 improved and extended [34] 27] 18] A number of different preemptive policies have been proposed and studied for scheduling parallel jobs in multiprocessors [33] 29] 38] 20] 28] 26] [23] [24] 2] In particular, experimental and simulation studies have shown that the DEQ algorithm yields low mean completion times under a variety of workloads and is reported to possess desirable properties of a good scheduler [33] 17] 16] DEQ was first introduced to parallel scheduling by ....

[Article contains additional citation context not shown here]

T. Nguyen and R. Vaswani and J. Zahorjan, Using Runtime Measured Workload Characteristics in Parallel Processor Scheduling, Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science, 1162: edited by D.G. Feitelson and L. Rudolph, Springer-Verlag, 1996, pp. 175--199.

....scheduling contexts [20, 32, 3] However, we believe that such strategies must also consider the efficiency with which parallel programs execute. Previous work has suggested approaches for obtaining estimates of a job s efficiency during runtime [3] Recently, Nguyen, Vaswani, and Zahorjan [25, 24] have experimentally demonstrated that such estimates can in fact be determined by monitoring the job s execution during run time. They also find that the performance of scheduling policies that use these estimates achieve performance surprisingly close to that possible when perfect a priori ....

.... since we are only comparing dynamic scheduling algorithms (i.e. we are not comparing with static policies) Furthermore, experimental studies have demonstrated that although overhead is required to reallocate processors, the overhead does not significantly impact dynamic scheduling algorithms [25, 24]. For the purposes of simplifying the implementation of our simulator, we also assume that the number of processors allocated to a job may be fractional. Such an assumption does not affect the qualitative results of this work. As well, the effects of this assumption on the quantitative results are ....

[Article contains additional citation context not shown here]

T. Nguyen, R. Vaswani, and J. Zahorjan. Using runtime measured workload characteristics in parallel processor scheduling. In Proceedings of the IPPS'96 Workshop on Job Scheduling Strategies for Parallel Processor Scheduling, pages 93--104, Waikiki, HI, April 1996.

....jobs are not multiples of each other, e.g. not all powers of two. An interesting option available to systems with dynamic partitioning is to evaluate the speedup characteristics of applications as they run, and use this information to determine the optimal partition size. This is done as follows [430, 428, 429]. When a job is submitted, the system executes it for short periods on different partition sizes. Hardware and software monitors are used to gauge the efficiency of the execution, by tabulating processor stalls and synchronization delays. The expected speedup is then computed as the product of the ....

T. D. Nguyen, R. Vaswani, and J. Zahorjan, "Using runtime measured workload characteristics in parallel processor scheduling ". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 155--174, Springer-Verlag, 1996. Lecture Notes in Computer Science Vol. 1162.

....An interesting benefit of malleable jobs is that the option for changes can be used to allow the system to collect information about the job at runtime, by trying several configurations and checking the resulting performance. This information can later be used to guide allocation decisions [37]. This approach has obvious advantages over requiring the information to be available in advance, as is needed for moldable jobs. 4 Assumptions About Permissible Actions A scheduler must execute in the environment of an existing operating system and machine architecture. This environment ....

T. D. Nguyen, R. Vaswani, and J. Zahorjan, "Using runtime measured workload characteristics in parallel processor scheduling ". In Job Scheduling Strategies for Parallel Processing II, D. G. Feitelson and L. Rudolph (eds.), Springer-Verlag, 1996. Lecture Notes in Computer Science.

.... which to gain knowledge about job resource requirements must be broadened to include: ffl consider user provided information (while recognizing that it is historically quite unreliable, in part because users aren t careful about making good estimates) ffl measure efficiency during execution [60], and increase processor allocations only for jobs that are using their currently allocated processors effectively; ffl keep track of execution time and speedup knowledge from past executions on a class by class basis, and use that information. All identifying characteristics associated with ....

....policy will penalize users who intentionally misestimate the characteristics of the jobs they submit. In addition to information available before the job is initiated, it has been demonstrated that it is feasible to measure some job characteristics (such as efficiency) while the job is executing [60]. Then, assuming the job is malleable, the system can take appropriate action with respect to giving additional processors, or taking some away from the job. Finally, if some historical information is retained, then observed behavior of previous jobs with certain characteristics can be used to ....

[Article contains additional citation context not shown here]

T. D. Nguyen, R. Vaswani, and J. Zahorjan, "Us- ing runtime measured workload characteristics in parallel processor scheduling ". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 155--174, Springer-Verlag, 1996. Lecture Notes in Computer Science Vol. 1162.

.... Recent work has highlighted the benefits of extracting information from applications during runtime [15] For example, Nguyen et al. have shown that extracting runtime information can be minimally intrusive and can substantially improve the per formance of a parallel job scheduler [27], whereas Gibbons proposed a simpler system to correlate runtimes to different job queues [16] In either case, information gathered from tasks as they run can support job scheduling and allocation. The Hector distributed runtime environment is intended to support this model [32] It uses a ....

T.D. Nguyen, R. Vaswani, and J. Zahorjan, Using Run--Time Measured Workload Characteristics in Parallel Processing Scheduling, IPPS '96 Workshop on Job Scheduling Strategies for Parallel Processing (1996).

....system uses a continuous stream of performance and availability information as its input. Parallel systems can benefit from gathering performance information at run time. For example, Nguyen, et al. have shown that adding runtime instrumenta tion can improve the performance of job scheduling [21]. Gibbons has shown that even rough esti mates of run time, gathered over many runs of programs, can help the scheduler approach the be havior of the well known shortest job first (SJF) algorithm [22] Harden, et al. have shown that a hybrid hardware software approach to monitoring can be ....

T. D. Nguyen, R. Vaswani, and J. Zahorjan, "Using Runtime Measured Workload Character- istics in Parallel Processing Scheduling", IPPS '96 Workshop on Job Scheduling Strategies for Parallel Processing, Honolulu, HI, April 1996. 8

....events implicit in parallel applications to estimate load imbalances. The local schedulers are able to determine from this data when to schedule parallel applications so that multiple processes of a job have a high probability of being scheduled simultaneously. Nguyen, Vaswani and Zahorjan [NVZ96b, NVZ96a] use a combination of code instrumentation and hardware monitors to determine run time characteristics of iterative applications. By varying the processor allocations over several iterations of a loop, the scheduler can determine application characteristics. Although Nguyen, et al. show the ....

T.D. Nguyen, R. Vaswani, and J. Zahorjan. Using runtime measured workload characteristics in parallel processor scheduling. In Proceedings of IPPS '96 Workshop on Job Scheduling Strategies for Parallel Processing, pages 93--104, April 1996.

....the degree of simi larity between clusters. H. Monitoring Resource Availability Parallel systems can benefit from gathering performance information at run time. For ex ample, Nguyen, et al. have shown that adding run time instrumentation can improve the performance of job scheduling [52]. Gibbons has shown that even rough estimates of run time, gathered over many runs of programs, can help the scheduler approach the behavior of the well known short est job first (SJF) algorithm [27] Harden, et al. have shown that a hybrid hardware software ap proach to monitoring can be ....

T. D. Nguyen, R. Vaswani, and J. Zahorjan, "Using Runtime Measured Workload Character- istics in Parallel Processing Scheduling", IPPS '96 Workshop on Job Scheduling Strategies for Parallel Processing, Honolulu, HI, April 1996.

....questions through three sets of measurements. First, we examine application speedup, and the sources of speedup loss. Our results confirm that there is considerable variation among jobs, and provide information that will support work on the use of speedup information in making scheduling decisions [13, 23]. Second, because it is at least burdensome, and perhaps impossible, to accurately collect and supply such information at job submission time, we look at the problem of estimating job speedup at runtime. We first demonstrate a technique for estimating instantaneous speedup at runtime that is ....

....much worse for compiler parallelized applications than for hand coded applications. Most speedup curves are relatively smooth and roughly convex shaped. This implies that speedup values for a relatively few allocations might allow reasonably accurate extrapolation to other allocations. See [23] for an application of this idea to scheduling. For most hand coded applications, there is an allocation beyond which they slow down gradually. With the exception of ARC2D when parallelized by KAP, all compiler parallelized jobs slow down significantly after achieving their peak speedups. ....

[Article contains additional citation context not shown here]

T. D. Nguyen, R. Vaswani, and J. Zahorjan. Using Runtime Measured Workload Characteristics in Parallel Processor Scheduling. In Proceedings of the IPPS'96 Workshop on Job Scheduling Strategies for Parallel Processing, Apr. 1996.

No context found.

Nguyen TD, Vaswani R, Zahorjan J. Using runtime measured workload characteristics in parallel processor scheduling. Science, vol. 1162). Springer: Berlin, 1996.

No context found.

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling", in JSSPP volume 1162 of Lectures Notes in Computer Science. Springer-Verlag, 1996. 32

No context found.

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling", in JSSPP volume 1162 of Lectures Notes in Computer Science. Springer-Verlag, 1996.

No context found.

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". JSSPP, vol. 1162 of Lectures Notes in Computer Science. Springer-Verlag, Univ. of Washington, 1996.

No context found.

T. D. Nguyen, J. Zahorjan, R. Vaswani, "Using Runtime Measured Workload Characteristics in Parallel Processors Scheduling". Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science, vol. 1162, Springer-Verlag, Univ. of Washington, 1996.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC