Vijay K. Naik, Sanjeev K. Setia, and Mark S. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. In Proceedings of Supercomputing '93, pages 824--833, November 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....theoretic model to investigate how parallel processing overheads cause efficiency to decrease with larger processor allocations. In a later study [75] they go on to show that dynamic partitioning of the system beats static partitioning at moderate and heavy loads. Naik, Setia and Squillante [56] show that dynamic partitioning allows much better performance than fixed partitioning, but that much of the difference in Actually, EASY only guarantees that the first job in the queue will not be delayed. performance can be obtained by using knowledge of job characteristics, and assigning ....

V. K. Naik, S. K. Setia, and M. S. Squillante, "Performance analysis of job scheduling policies in parallel supercomputing environments". In Supercomputing '93, pp. 824--833, Nov 1993.

....processors in the system are not divided before the computation. When a new job arrives, a job manager in the system first locates idle processors and then allocates certain number of those idle processors to that job according to some processor allocation policies, e.g. those described in [2, 10, 14, 15,17, 18, 20]. Therefore, the boundary lines are drawn during the computation and will disappear after the job is terminated. Normally the static partitioning is used for very large systems, while the adaptive partitioning is adopted in systems, or subsystems of small to medium size. One disadvantage of space ....

V. K. Naik, S. K. Setia and M. S. Squillante, Performance analysis of job scheduling policies in parallel supercomputing environments, Proceedings of Supercomputing '93, Nov. 1993, pp.824-833.

.... But what if the J jobs are not known to be identical in fact, what if nothing is known about them In that case, most papers from academia suggest dynamic space slicing: jobs are space sliced, and as soon as a job terminates the remaining jobs are re allocated to make use of the free PEs [NSS93, R 94, Sev89, ST93, EZL89, SHCV94, LV90, MVZ93, MZ95, TG89, D 95] Unfortunately, in most cases, this result is obtained by facilely neglecting the exorbitant cost of re allocating jobs (see Section 3.7.4) Industry practices gang scheduling: On the other hand, the industry has always ....

V.K. Naik, S.K. Setia, and M.S. Squillante. Performance analysis of job scheduling policies in parallel supercomputing environments. In Supercomputing, pages 824--833, November 1993.

....with EQS even when repartitioning is assumed to have zero cost. This is an impressive result since FB PWS commits to a processor allocation at job arrival time. Previous static allocation policies have generally not been competitive with dynamic policies such as EQS under zero repartitioning cost [13, 29, 19, 4]. They also show that FB PWS and FB ASP can substantially outperform EQS under an ad hoc model of repartitioning costs that is intended to illustrate the possible impact of repartitioning overheads on relative policy performance. In this paper, we further investigate the relative performance of ....

V.Naik, S. Setia, and M. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. Proceedings of Supercomputing '93, November 1993.

....active on any given processor. Second, jobs generally make more efficient use of the processing resources given smaller processors allocations. As a result, providing the scheduler with some flexibility in allocating processors can significantly improve overall performance [GST91,Sev94,NSS93,RSD 94] In most systems, users specify precisely the number of processors which should be allocated to each job, a practice that is known as rigid scheduling. In adaptive scheduling disciplines, the user specifies a minimum processor allocation, usually resulting from constraints due to ....

Vijay K. Naik, Sanjeev K. Setia, and Mark S. Squillante. Performance analysis of job scheduling policies in parallel supercomputing environments. In Proceedings of Supercomputing '93, pages 824--833, 1993.

.... available processors into equal sized partitions has been studied with analytic models [Sev89, MEB91, ST93, SST93] compared with adaptive policies [Sev89, MEB91, GST91, RSD 94, ST93] compared with timeslicing policies [SST93] and compared with both adaptive and dynamic policies [NSS93b, NSS93a, NSS97] The general conclusion is that fixed partitioning can perform well if the workload is wellunderstood and jobs tend to have low variation in parallelism and service times. In many cases, a particular size of fixed partitioning outperforms all other policies for a particular system load ....

....system wide minimum. This policy performed nearly as well as fixed policies at particular loads, but was more consistent across different configurations. A dynamic policy allowing limited pre emption was also studied, but it tended to perform poorly at high load. Further study by the same authors [NSS93a] showed that the dynamic policy handled transient bursts of arrivals much better than the fixed policies. Chiang et al. CMV94] compare adaptive policies that use application characteristics such as average parallelism and pws with various versions of ASP. They find that for RTC policies to ....

Vijay K. Naik, Sanjeev K. Setia, and Mark S. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. In Proceedings of Supercomputing '93, pages 824--833, November 1993.

....The test run along with assuming actual runtime is 10 of the requested run time during backfill, improves average slowdown by 50 for a job trace from a 430 node cluster. Previous papers that compare spatial equipartitioning to nonpreemptive policies either do not consider memory requirements [15, 17], or assume that each job has a uniform distribution for the minimum number of processors it can execute on to represent the job s memory requirement ( 16, 19] 3. NCSA Workload Characterization This section provides an overview of the O2K batch workloads used to evaluate the scheduling ....

V. K. Naik, S. K. Setia, and M. S. Squillante. Performance analysis of job scheduling policies in parallel supercomputing environments. Proc.

....can improve overall system utilization and can allocate each process a fair share of the systems resources, this approach can lead to significant performance degradation for parallel applications when executed simultaneously with other sequential or parallel applications. A number of studies [16, 32, 35, 47] have shown that parallel programs often do not perform well in multiprogrammed environments due to the following reasons: ffl Context switching. If there are more processes in the system than the actual number of physical processors, the execution of the processes will be multiplexed using the ....

....parallel applications time share processor resources with other parallel or sequential applications. The high overheads associated with context switching, together with poor cache utilization, can result in poor parallel execution performance in time shared multiprogrammed environments [16, 32, 35, 47]. For programs with dynamically varying behavior, such as programs where the number of loop iterations changes dynamically across invocations, it is not possible to determine at compile time the number of processors that can be used most efficiently at each invocation. Furthermore, if the target ....

V. Naik, S. Setia, and M. Squillante, Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments, Supercomputing, 1993, pp. 824-833.

....may have to time share processor resources with other parallel or sequential applications. Due to the high overheads encountered when context1 switching, due to poor cache utilization, and due to other related problems, parallel programs often do not perform well in multiprogrammed environments [5, 8, 9, 14]. A given number of processors must be dedicated to the parallel application to obtain the best overall performance. However, since the system load changes dynamically, the number of processors that can be dedicated to a parallel application in a given time period also changes. To avoid ....

V. Naik, S. Setia, and M. Squillante, Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments, In Supercomputing, 1993, pp. 824-833.

....applications may have to time share processor resources with other parallel or sequential applications. Due to the high overheads encountered when context switching, poor cache utilization, and other related problems, parallel programs often do not perform well in multiprogrammed environments [4, 7, 8, 13]. Furthermore, since the system load changes dynamically, the number of processors that can be dedicated to a parallel application in a given time period also changes. We propose a comprehensive dynamic processor allocation scheme for shared memory multiprocessor systems with a multiprogrammed ....

V. Naik, S. Setia, and M. Squillante, Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments, Supercomputing, 1993, pp. 824-833.

.... perfectly OK to preempt PEs from a batch job in favor of an interactive one, but it is undesirable to do so the other way round [422, 30] A possible mechanism to guarantee a certain level of service is to prescribe a minimal partition size, and possibly different sizes for different job classes [421]. A similar consideration applies when the partitions are assigned to phases of jobs, rather than to complete jobs. This allows a job to relinquish a large partition when it enters a serial phase in its computation. However, in order to avoid penalizing such jobs, the serial phase (i.e. the ....

....schemes suffer from the same memory oblivion as adaptive partitioning. As a result, jobs may find that they have a (small) number of PEs, but cannot use them effectively because they do not have enough memory [452, 515] This situation is improved if minimal partition 41 sizes are imposed [421]. 3.5 Processors Allocated Singly Dynamically changing partitions can also be created by the individual allocation of single PEs. The fact that a number of PEs end up executing threads belonging to the same application, and thus implicitly defining a partition allocated to that application, is ....

[Article contains additional citation context not shown here]

V. K. Naik, S. K. Setia, and M. S. Squillante, "Performance analysis of job scheduling policies in parallel supercomputing environments". In Supercomputing '93, pp. 824--833, Nov 1993.

....that among dynamic policies that do not use job demand information, policies that allocate processing power approximately equally among executing jobs (i. e, EQ policies) have high performance when the coefficient of variation, C D , in cumulative job processing requirement is greater than one [LV90, ST91, NSS93a, NSS93b, MVZ93, MV93b]. Also, several studies have shown that particular RTC policies have poorer performance than particular dynamic policies [ZM90, MEB91, ST91, NSS93b] As a third example, there is consensus in the literature that RTC policies have higher performance if they allocate fewer processors to a job as ....

.... policies) have high performance when the coefficient of variation, C D , in cumulative job processing requirement is greater than one [LV90, ST91, NSS93a, NSS93b, MVZ93, MV93b] Also, several studies have shown that particular RTC policies have poorer performance than particular dynamic policies [ZM90, MEB91, ST91, NSS93b]. As a third example, there is consensus in the literature that RTC policies have higher performance if they allocate fewer processors to a job as (instantaneous or average) system load increases [Se89, GST91, NSS93a, NSS93b, ST93] The data in [ZM90, MEB91, RSDS93, SRDS93] also support this ....

[Article contains additional citation context not shown here]

V.Naik, S. Setia, and M. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. Proc. of Supercomputing'93, November 1993.

.... perfectly OK to preempt PEs from a batch job in favor of an interactive one, but it is undesirable to do so the other way round [260, 19] A possible mechanism to guarantee a certain level of service is to prescribe a minimal partition size, and possibly different sizes for different job classes [259]. Efficiency can also become an issue. Consider a job that is written so that the work is divided into 8 equal pieces (e.g. 8 chores) Running such a job on a partition of 7 PEs would not accrue any benefits over a partition of only 4 PEs, leading to waste of resources [239] A possible solution ....

....more detail in Section 7.2.2. A variant based on feedback has also been proposed [311] 6 Implications of the Sharing Scheme A number of studies have compared the effectiveness of different scheduling schemes, typically using their impact on the performance of specific applications as a metric [225, 230, 217, 152, 71, 122, 259, 57]. This section summarizes these results and adds many additional considerations, such as the support for different programming models and the influence on user satisfaction. These considerations are divided into four areas: the interaction between scheduling and the application, the impact on ....

[Article contains additional citation context not shown here]

V. K. Naik, S. K. Setia, and M. S. Squillante, "Performance analysis of job scheduling policies in parallel supercomputing environments". In Supercomputing '93, pp. 824--833, Nov 1993.

....Foundation under grants CCR 9024144 and CDA 9024618. 1 1 Introduction The increasing use of parallel processor systems has led to the development of a number of multiprogrammed processor allocation policies. Many studies have compared the performance of specific processor allocation policies [7, 13, 14, 15, 16, 20, 21, 22, 25, 33, 34, 35, 37, 39, 41], which has led to a diverse set of results concerning relative policy performance over numerous specific regions of the workload parameter space. 1 For example, ffl The Adaptive Static Partitioning (ASP) policy has been shown to have higher performance than several static allocation policies, ....

.... and highly correlated workloads when CD is moderate to high [14] ffl The spatial equipartitioning (EQS) policy is also shown to have high performance for specific measurement workloads [39, 20, 8] and a particular workload that consists of a mixture of application types and has high CD [21, 22]. These results show particular policies to perform well over narrow regions of the parameter space but it is not clear whether or how the various results generalize. In particular, it is not clear which workload parameters determine relative policy performance or how the policies compare for say ....

[Article contains additional citation context not shown here]

V.Naik, S. Setia, and M. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. Proceedings of Supercomputing'93, November 1993.

....Our baseline strategy is AVG, which assigns free processors to queued jobs in first come first served order, giving each job no more than A processors, where A is the average parallelism of the job. Several studies have shown that this strategy performs well for a range of workloads [21] 11] [15] [22] 4] 6] An important feature of this strategy is that it is work conserving: if even one processor is free, the job at the head of the queue will be forced to begin execution immediately. From the system s point of view, work conservation is expected to yield high utilization, but from ....

Vijay K. Naik, Sanjeev K. Setia, and Mark S. Squillante. Performance analysis of job scheduling policies in parallel supercomputing environments. In Supercomputing '93 Conference Proceedings, pages 824--833, March 1993.

....benefit of doing so. Our baseline strategy is AVG, which assigns free processors to queued jobs in FIFO order, giving each job no more than A processors, where A is the average parallelism of the job. Several studies have shown that this strategy performs well for a range of workloads [17] 9] [12] [19] 1] 2] The problem with this strategy is that it forces users to accept decisions that are contrary to their interests. For example, if there is a large job at the head of the queue, it will be forced to run on any available cluster, even a single processor. From the system s point of ....

Vijay K. Naik, Sanjeev K. Setia, and Mark S. Squillante. Performance analysis of job scheduling policies in parallel supercomputing environments. In Supercomputing '93 Conference Proceedings, pages 824--833, March 1993.

.... show that dynamic partitioning allows much better performance than fixed partitioning, but that much of the difference in performance can be obtained by using knowledge of job characteristics, and assigning non preemptive priori10 ties to certain job classes for admission to fixed partitions [57]. McCann and Zahorjan found that efficiencypreserving scheduling using folding allowed performance to remain much better than with equipartitioning (EQUI) as load increases [52] Padhye and Dowdy compare the effectiveness of treating jobs as moldable to that of exploiting their malleability by ....

V. K. Naik, S. K. Setia, and M. S. Squillante, "Performance analysis of job scheduling policies in parallel supercomputing environments". In Supercomputing '93, pp. 824--833, Nov 1993.

....has long been an active area of research. The problem is typically decomposed into two steps: allocation, where individual processes are placed on processors, and dispatching, where those processes are scheduled over time. The allocation step for parallel jobs has been investigated in detail [31, 33, 66, 70, 101, 118, 119, 120, 126, 139, 149]. Given the popular single programmultiple data (SPMD) parallel programming model, it has generally been found that the response time and throughput of the workload are best when competing processes from different jobs can share the same processor [35, 53, 82, 105, 144, 148, 173] In this work, ....

....can be used. Pure space sharing techniques that achieve good response time and throughput require that applications are malleable to the number of available processors, a non trivial programming task. A large number of studies have focused on the allocation step of parallel job scheduling [31, 33, 66, 70, 101, 118, 119, 120, 126, 139, 149]. In this dissertation, we focus on the second step of time sharing processes over time. Two popular methods exist for timesharing competing processes: local scheduling and explicit coscheduling. When processes are locally scheduled by the operating system on each workstation, frequently ....

Vijay K. Naik, Sanjeev K. Setia, and Mark S. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. In Proceedings of Supercomputing '93, pages 824--833, November 1993.

....is assigned one of the partitions. With adaptive non preemptive policies, the partition size assigned to an application is computed before the application starts execution and may vary in successive application executions depending upon the current state of the system and possibly other parameters [Sevc89, ZM90, NSS93, Sevc94, RSDSC94]. With preemptive policies, an executing application may be interrupted and part of its processors may be assigned to another application. Time sharing policies, by definition, are preemptive as tasks of various applications may be assigned to each processor. Processors may service their tasks in ....

V.K. Naik, S.K. Setia, M.S. Squillante, "Performance analysis of job scheduling policies in parallel supercomputing environments," Proc. of Supercomputing 93, 1993, pp. 824-833.

....employed in a number of commercial systems. This is due in part to its low system overhead and its simplicity from both the system and application viewpoints. The static scheduling approach, however, can lead to relatively low system throughputs and resource utilizations under nonuniform workloads [34, 21, 22, 25, 35], as is common in scientific engineering computing environments [6] Adaptive partitioning policies, where the number of processors allocated to a job is determined when jobs arrive and depart based on the current system state, have also been considered in a number of research studies [14, 42, 8, ....

.... 22, 25, 35] as is common in scientific engineering computing environments [6] Adaptive partitioning policies, where the number of processors allocated to a job is determined when jobs arrive and depart based on the current system state, have also been considered in a number of research studies [14, 42, 8, 21, 22, 33, 3, 25]. This approach tends to outperform its static counterparts by adapting partition sizes to the current load. However, the performance benefits of adaptive partitioning can be limited due to its inability to adjust scheduling decisions in response to subsequent workload changes. These potential ....

[Article contains additional citation context not shown here]

V. K. Naik, S. K. Setia, and M. S. Squillante. Performance analysis of job scheduling policies in parallel supercomputing environments. In Proc. Supercomputing '93, 824--833, 1993.

....employed in a number of commercial systems. This is due in part to its low system overhead and its simplicity from both the system and application viewpoints. The static scheduling approach, however, can lead to relatively low system throughputs and resource utilizations under nonuniform workloads [33, 26, 27, 30, 34], which can be common in scientific and engineering computing environments. Adaptive partitioning policies, where the number of nodes allocated to a job is determined when jobs enter and leave based on the current system state, have also been considered in a number of research studies [21, 38, 16, ....

....27, 30, 34] which can be common in scientific and engineering computing environments. Adaptive partitioning policies, where the number of nodes allocated to a job is determined when jobs enter and leave based on the current system state, have also been considered in a number of research studies [21, 38, 16, 26, 27, 32, 7, 30]. This approach tends to outperform its static counterparts by adapting partition sizes to the current load. On the other hand, the performance benefits of adaptive partitioning can be limited due to its inability to adjust scheduling decisions in response to subsequent workload changes. These ....

[Article contains additional citation context not shown here]

V. K. Naik, S. K. Setia, and M. S. Squillante. Performance analysis of job scheduling policies in parallel supercomputing environments. In Proceedings of Supercomputing '93, pages 824--833, November 1993.

....the needs of both interactive applications belonging to workstation owners and multiple batch applications trying to scavenge idle cycles from the cluster. One of the scheduling policies that has been shown to have good performance in dedicated parallel environments is dynamic space sharing [6, 13, 17]. Under dynamic space sharing, the processors of a parallel system are divided into disjoint partitions that are allocated to individual jobs. However, the number of processors allocated to a partition can be changed dynamically in response to events such as new job arrivals or job departures. In ....

....applications, and presents measurement results for the cost of dynamic reconfiguration. Section 5 describes future work and conclusions. 2 Motivation for Dynamic Spacesharing In the last few years, several studies have examined job scheduling strategies for dedicated parallel environments [6, 13, 11, 7, 17]. Policies that employ dynamic space sharing have been shown to outperform other policies such as static space sharing and gangscheduling with static partitioning for many workloads. In addition to its advantages in dedicated environments, dynamic space sharing has some natural advantages in ....

[Article contains additional citation context not shown here]

V. K. Naik, S. K. Setia, and M. S. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. In SuperComputing '93, November 1993.

No context found.

Vijay K. Naik, Sanjeev K. Setia, and Mark S. Squillante. Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments. In Proceedings of Supercomputing '93, pages 824--833, November 1993.

No context found.

V. K. Naik, S. K. Setia, and M. S. Squillante, "Performance analysis of job scheduling policies in parallel supercomputing environments," in Supercomputing '93, November 1993, pp. 824--833.

No context found.

V. Naik, S. Setia, and M. Squillante, Performance Analysis of Job Scheduling Policies in Parallel Supercomputing Environments, Supercomputing, 1993, pp. 824-833.

First 50 documents  Next 50

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