C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....can be achieved with few reassignments of processors. Of course, much more research is required to make this theoretical understanding useful in a practical setting. In fact, performance in such systems has been already studied using simulation, experimental, and queuing theoretical approaches [2, 13, 17, 18, 24, 27, 1]. In this perspective, our research constitutes a theoretical confirmation of these efforts. 1.1 Modeling Job Execution In our model, all jobs arrive at time zero. That is, we adopt a batch job processing model. It would be more general to allow jobs to arrive at arbitrary times. However, this ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.

....) This gives ffi ffiT , which is the length of time for EQUI s to complete the new fully parallelizable phase with work fi ffiT . 11 Open Problems The performance of Equi partition has been studied extensively using simulation, experimental, and queuing theoretical approaches [3, 13, 16, 17, 24, 28, 2]. Our research constitutes a theoretical confirmation of these efforts. The main open problem is to close the gaps between the lower bounds on the competitive ratio known for general non clairvoyant schedulers and those known for the specific schedulers Equi partition and Balance. This gap is ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.

....n) However, the batch model assumes that the number of jobs n is at most the number of processors p because all these jobs arrive and are executed at once. Because our jobs arrive at arbitrary times, n is assumed to be much bigger then p. 28 13 Open Problems and queuing theoretical approaches [3, 13, 16, 17, 24, 28, 2]. Our research constitutes a theoretical confirmation of these efforts. The main open problem is to close the gaps between the lower bounds on the competitive ratio known for general non clairvoyant schedulers and those known for the specific schedulers Equi partition and Balance. This gap is ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.

....to approach ideal (i.e. no overhead) EQUI, and Anastasiadis et al. 3] show that, by setting the processor allocation of moldable jobs based on some known job characteristics, disciplines with little or no preemption can do nearly as well as EQUI. Other Factors in Scheduling McCann and Zahorjan [52] studied the scheduling problem where each job has a minimum processor allocation due to its memory requirement. They find that a discipline based on allocation by a buddy system consistently does well. Alverson et al. 2] describe the scheduling policy for the Tera MTA, which includes ....

C. McCann and J. Zahorjan, "Scheduling memory constrained jobs on distributed memory parallel computers". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 208--219, May 1995.

....3) Other Factors in Scheduling Much of the research available in the literature focuses on scheduling processors. In practice, however, other resources (e.g. disk, memory, communication infrastructure) may also need to be considered. There is some research in multi resource scheduling [6] 15] [70], but it still remains a largely unexplored issue. 7.1.2. Application Scheduling Application schedulers are an essential component of scheduling solutions for computational grids. Computational grids are platforms for the execution of parallel 105 jobs that are composed by geographically ....

C. McCann and J. Zahorjan. Scheduling Memory Constrained Jobs on Distributed -Memory Parallel Computers. In SIGMETRICS Conference on Measurement and Moldeling of Computer Systems, pp. 208-291, May 1995.

.... 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 leaned towards gang scheduling. Most vendors already ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of the ACM SIGMETRICS and Performance Joint Conference on Measurement and Modeling of Computer Systems, pages 208--219, May 1995.

....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 policies in Section 4, with particular attention to the job characteristics that differ from system workloads reported previously in [6, 9, 4, 5, 22] Table 4 ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. Proc. ACM SIGMETRICS Conf. on Measurement and Modeling of Computer Systems, Ottawa, May 1995, pp. 208-219.

....can be achieved with few reassignments of processors. Of course, much more research is required to make this theoretical understanding useful in a practical setting. In fact, performance in such systems has been already studied using simulation, experimental, and queuing theoretical approaches [2, 13, 17, 18, 24, 27, 1]. In this perspective, our research constitutes a theoretical con rmation of these e orts. 1.1 Modeling Job Execution In our model, all jobs arrive at time zero. That is, we adopt a batch job processing model. It would be more general to allow jobs to arrive at arbitrary times. However, this ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208-219, 1995.

.... This result provides a theoretical foundation for analyzing and understanding the performance of the DEQ policy, which, along with its various derivatives, has been shown to be superior to nonpreemptive algorithms in recent simulation and experimental studies [33] 17] 38] 16] 20] [21]. The remainder of the paper is organized as follows. We complete section 1 with a further description of the problem and a discussion of related work. In section 2, we give a formal definition of the DEQ allocation policy. Then we establish a lower bound on the optimal total completion time for ....

C. McCann and J. Zahorjan, Scheduling memory constrained jobs on distributed memory parallel computers, in Proceedings of the International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, ACM, New York, 1995, pp. 208--219.

....of multiple resource requirements and general load balancing methods for homogeneous parallel processing systems. Recent research in job scheduling for a single server has demonstrated the benefits of including information about the memory requirements of a job in addition to its CPU requirements [13, 14]. The generalized # resource single server scheduling problem was studied in [10] where it was shown that simple backfill algorithms based on multi dimensional packing heuristics consistently outperform single resource algorithms, with increasing #.These efforts all suggest that the local JMS at ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory computers. In Proc. ACM SIGMETRICS Joint Intl. Conference on Measurement and Modeling of Computer Systems, pages 208--219, 1996.

....establishments, resource management of these machines becomes a more pressing issue. Job scheduling is an important aspect of this work, and much research on the topic of job scheduling for parallel computers has been done in recent years. Some have focused on distributed memory machines [7, 8, 9, 11, 14], others on more general systems [3, 6, 10, 15] Many different strategies (static vs. dynamic, time sharing vs. space sharing, etc. have been investigated and compared. All these studies make underlying assumptions about the workload in order to quantify their results. Some of these studies use ....

C. McCann and J. Zahorjan. "Scheduling Memory Constrained Jobs in Distributed Memory Parallel Computers". In Proceedings of ACM SIGMETRICS Conference, 1995. 23(1): p. 208-219.

....parallelism. This result provides a theoretical foundation for analyzing and understanding the performance of the DEQ policy which, along with its various derivatives, has been shown to be superior to non preemptive algorithms in recent simulation and experimental studies [33] 17] 38] 16] 20] [21]. The remainder of the paper is organized as follows: We complete Section 1 with a further description of the problem and a discussion of related work. In Section 2, we give a formal definition of the DEQ allocation policy. Then we establish a lower bound on the optimal total completion time for ....

C. McCann and J. Zahorjan, Scheduling Memory Constrained Jobs on Distributed Memory Parallel Computers, Proceedings of International Joint Conference on Measurement & Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, 1995, pp. 208--219.

....upon job arrivals and departures [33, 13] The advantage of this dynamic scheduling policy is that it may adjust to a job s changing demands for processors over time. Recent studies examine scheduling algorithms that consider another important characteristic, the memory requirements of jobs [37, 4, 29, 23, 30, 1, 27, 28]. These studies identify memory (as well as processors) as being a critical resource for the effective execution of parallel programs and devise and investigate scheduling policies that ensure minimum processor allocations for jobs in order to secure sufficient memory resources. Our work does not ....

....a specified number of processors, p i . This models the efficiency of the parallel job. A number of models of parallel system and parallel program performance have been proposed and studied [10, 35, 7, 32] We use the following execution rate function, used in a number of previous studies [5, 23, 30], which has been derived from an execution rate function (also called an execution signature) proposed by Dowdy [7] F = S(p i ) 1 fi i ) p i fi i p i : In this equation S(p i ) is the speedup obtained when the job is executed on p i processors and fi i is the parameter that is used to ....

[Article contains additional citation context not shown here]

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of the 1995 ACM SIGMETRICS Joint International Conference on Measurement and Modeling of Computer Systems, pages 208--219, Ottawa, ON, May 1995. 23

....for distributed memory machines, where the page must be collocated with the thread that accesses it, or for UMA shared memory machines, where all memory is equally accessible. Regrettably, the interaction of memory management with scheduling in parallel systems has received only scant attention [452, 88, 396, 106, 515, 448, 447, 397, 550]. This interaction has great importance. Systems that use non preemptive partitioning typically do not provide any virtual memory or paging, because they cannot afford the overhead of idling a PE while waiting for a page fault to be serviced [452, 515, 256] All the memory is dedicated to a single ....

C. McCann and J. Zahorjan, "Scheduling memory constrained jobs on distributed memory parallel computers". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 208--219, May 1995.

.... has been done in the context of NUMA architectures on the automatic migration of pages from one memory to another, so as to improve locality of reference [42, 40, 302, 212] Regrettably, the interaction of memory management with scheduling in parallel systems has received only scant attention [278, 49, 239, 57, 310, 274, 273, 240, 331]. This interaction has great importance. Systems that use non preemptive partitioning typically do not provide any virtual memory or paging, because they cannot afford the overhead of idling a PE 17 It is often said that users might also create spurious threads that do not perform any useful ....

C. McCann and J. Zahorjan, "Scheduling memory constrained jobs on distributed memory parallel computers". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 208--219, May 1995.

.... Slicing and Memory Pressure Previous work about incorporating memory considerations into scheduling algorithms has been quite limited, and included ideas such as the following: In systems that use space slicing, place a lower bound on partition size so that enough memory will be available [16,17,13]. When the partition size is adjustable, do not reduce it too much, because small partitions cause jobs to run longer and thus increase the memory pressure [15] In systems that use swapping, make the residence time proportional to the memory footprint size in order to amortize the cost of ....

C. McCann and J. Zahorjan, "Scheduling memory constrained jobs on distributed memory parallel computers". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 208--219, May 1995.

....slicing, because chances are that a new job will have a short runtime [24,23] As production workloads do indeed exhibit a high variability [6] it follows that gang scheduling will reduce mean response time. Indeed, gang scheduling has even been advocated in conjunction with dynamic partitioning [21]. 3 Simulation Results 3.1 The Compared Scheduling Schemes In order to demonstrate the ideas described above, we simulate the performance of a multicomputer subjected to a realistic workload and using one of a set of different scheduling schemes. these are: FCFS: the base case we use for ....

C. McCann and J. Zahorjan, "Scheduling memory constrained jobs on distributed memory parallel computers". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 208--219, May 1995.

....One main reason is that the allocations are made for the lifetime of a job. Dynamic spacesharing policies eliminate this problem. The idea behind dynamic policies is that idle processors should be taken away from a job and allocated to another job that can fruitfully utilize the processor cycles [8]. One way to reduce the overhead (associated with taking a processor away from a job and allocating it to another job) in distributed memory systems is to wait until the computation reaches a desired point such that the overhead involved in taking a processor away is small. This implies that the ....

C. McCann and J. Zahorjan, "Scheduling Memory Constrained Jobs on Distributed Memory Parallel Computers," Proc. ACM Sigmetrics Conf., Ottawa, Canada, 1995, pp. 208-219.

....can be achieved with few reassignments of processors. Of course, much more research is required to make this theoretical understanding useful in a practical setting. In fact, performance in such systems has been already studied using simulation, experimental, and queuing theoretical approaches [1, 8, 9, 12, 13, 14, 18, 21]. In this perspective, our research constitutes a theoretical confirmation of these efforts. 1.1 Modeling Job Execution In our model, all jobs arrive at time zero (batch) It would be more general to allow jobs to arrive at anytime. However, this makes the scheduling problem much more difficult ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.

....moldable jobs based on some known job characteristics, disciplines with little or no preemption can do nearly as well as EQUI [3] 2.2. 8 Other factors in scheduling McCann and Zahorjan studied the scheduling problem where each job has a minimum processor allocation due to its memory requirement [53]. They find that a discipline based on allocation by a buddy system consistently does well. Alverson et al. describe the scheduling policy for the Tera MTA, which includes consideration of memory requirements [2] Brecht has carried out an experimental evaluation of scheduling in systems where ....

C. McCann and J. Zahorjan, "Scheduling memory constrained jobs on distributed memory parallel computers". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 208-- 219, May 1995.

....a specific adaptive partitioning policy, taking into account factors such as fragmentation overhead. Secondly, our study also differs in that we take into account the impact of synchronization between threads on the overall paging overhead incurred by a parallel application. McCann and Zahorjan [13] propose and evaluate several scheduling policies that take into account the constraints imposed by the memory requirements of parallel applications. Our work differs in that we consider memory constrained adaptive partitioning policies, whereas they assume a dynamic scheduling discipline. ....

....These assumptions result in an average job demand (considering both classes) of 1125 seconds, with the coefficient of variation of job demand (CD ) of 2.05. The speedup function used in our simulations is given by S(p) 1 fi)p= fi p) This speedup function has been used by several studies [13, 3, 5] and is shown in Figure 1. For a given number of processors (p) and given job demand (on one processor) the speedup function is used to compute the processing requirement of the job on p processors. In our simulations, we assume that fi is uniformly distributed between 30 and 300, the range ....

[Article contains additional citation context not shown here]

Cathy McCann and John Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. Technical Report UW-CSE-94-10-05, University of Washington, Department of Computer Science, 1994.

....space sharing [32] and gang scheduling [12] and several issues (e.g. space sharing vs time sharing [20] static vs dynamic partitioning [32] have been analyzed in the literature. From these studies a consensus has emerged that (i) it is necessary to coschedule the processes of a parallel job [10, 21], and (ii) allocating a equal share of the available processing capacity to competing jobs results in good performance [21, 4] To achieve these goals, a scheduling policy has to employ both timeand space sharing. The use of space sharing ensures that the processes of a job are coscheduled, while ....

....partitioning [32] have been analyzed in the literature. From these studies a consensus has emerged that (i) it is necessary to coschedule the processes of a parallel job [10, 21] and (ii) allocating a equal share of the available processing capacity to competing jobs results in good performance [21, 4]. To achieve these goals, a scheduling policy has to employ both timeand space sharing. The use of space sharing ensures that the processes of a job are coscheduled, while time sharing helps to achieve an equal division of the processing capacity among the jobs in the system. One of the policies ....

[Article contains additional citation context not shown here]

Cathy McCann and John Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of 1995 ACM Sigmetrics Conference, pages 208--219, May 1995.

....preemptive policies allow executing programs to be interrupted and dynamically reallocated a larger or smaller set of processors. Examples of preemptive policies include gang scheduling [17] 4] 6] time sharing [11] 25] 9] and dynamic space sharing [9] 18] 3] 26] 12] 13] [15]. Preemptive policies are optimal from an allocation point of view, since they allow for better resource utilization and can adapt to sudden changes in the workload intensity. However, the complexity of the run time environment for their implementation on an actual system and the overhead of ....

C. McCann and J. Zahorjan, "Scheduling Memory Constrained Jobs on Distributed Memory Parallel Computers," ACM SIGMETRICS, pp. 208-219, 1995.

....p because all these jobs arrive and are executed at once. Because our jobs arrive at arbitrary times, n is assumed to be much bigger then p. 13 Open Problems The performance of Equi partition has been studied extensively using simulation, experimental, and queuing theoretical approaches [3, 13, 16, 17, 24, 28, 2]. Our research constitutes a theoretical confirmation of these efforts. The main open problem is to close the gaps between the lower bounds on the competitive ratio known for general non clairvoyant schedulers and those known for the specific schedulers Equi partition and Balance. This gap is ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.

.... Parallel scheduling is composed of at least two interdependent steps: the allocation of processes to processors (space sharing) and the scheduling of the processes over time (time sharing) A large number of studies have focused on the processor allocation step of parallel job scheduling [5, 7, 17, 19, 27, 29, 30, 31, 33, 36, 37]; fewer have investigated the second step. We believe that a mixed approach, utilizing both space sharing and time sharing, is required to maintain interactive response times and high throughput. In this paper, we consider the problem after jobs have been allocated to a set of processors and focus ....

C. McCann and J. Zahorjan. Scheduling Memory Constrained Jobs on Distributed Memory Parallel Computers. In Proceedings of ACM SIGMETRICS'95/PERFORMANCE'95 Joint International Conference on Measurement and Modeling of Computer Systems, pages 208--219, May 1995.

....slicing, because chances are that a new job will have a short runtime [24, 23] As production workloads do indeed exhibit a high variability [6] it follows that gang scheduling will reduce mean response time. Indeed, gang scheduling has even been advocated in conjunction with dynamic partitioning [21]. 3 Simulation Results 3.1 The Compared Scheduling Schemes In order to demonstrate the ideas described above, we simulate the performance of a multicomputer subjected to a realistic workload and using one of a set of different scheduling schemes. these are: FCFS: the base case we use for ....

C. McCann and J. Zahorjan, "Scheduling memory constrained jobs on distributed memory parallel computers". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 208-- 219, May 1995.

.... combined with the fact that obtaining perfect knowledge is difficult, has lead to several efficient heuristics being proposed that make use of whatever limited or approximate information is available [6, 2] Most recently, other resource requirements of parallel jobs have been taken into account [12, 20, 17, 16]. In particular, it has been found that the memory demands of jobs, which are often non trivial in scientific computations, can greatly influence the scheduling decision if used in conjunction with speedup information. This issue is one of the major topics investigated in this thesis. Recent ....

Cathy McCann and John Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. Technical Report 94-10-05, University of Washington, October 1994.

....can be achieved with few reassignments of processors. Of course, much more research is required to make this theoretical understanding useful in a practical setting. In fact, performance in such systems has been already studied using simulation, experimental, and queuing theoretical approaches [2, 9, 11, 12, 18, 21, 1]. In this perspective, our research constitutes a theoretical confirmation of these efforts. 1.1 Modeling Job Execution In our model, all jobs arrive at time zero. That is, we adopt a batch job processing model. It would be more general to allow jobs to arrive at arbitrary times. However, this ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.

....However we point out these results may be generalized to other similar models. Our results are based on the DEQ scheduling policy 1 which, along with various derivatives, has been extensively studied in simulation, experimental and queuing theoretical approaches [27] 17] 33] 15] 21] 18] 3] [22]. Our analysis extends previous results to general DAGs and provides some theoretical explanation for the fact that DEQ performs well on jobs beyond those described by their parallelism profiles (discussed in [5] 1.1 The Machine Model As mentioned above, we distinguish two extreme cases of ....

C. McCann and J. Zahorjan, "Scheduling Memory Constrained Jobs on Distributed Memory Parallel Computers", Proceedings of International Joint Conference on Measurement & Modeling of Computer Systems, ACM Sigmetrics 95 and Performance 95, pp. 208-219, 1995.

....different and rather difficult proof. This provides a theoretical foundation for analyzing and understanding the performance of the DEQ policy which, along with its various derivatives, has been shown to be superior in recent simulation and experimental studies [24] 12] 29] 10] 16] 13] 2] [17]. Furthermore, our results can be extended to jobs which may change the number of processors required during their execution, including interactive jobs, which may block and therefore, need not be assigned to a CPU while waiting for user input. We also show that DEQ is robust in presence of faulty ....

C. McCann and J. Zahorjan, "Scheduling Memory Constrained Jobs on Distributed Memory Parallel Computers", Proceedings of International Joint Conference on Measurement & Modeling of Computer Systems, ACM Sigmetrics 95 and Performance 95, pp. 208-219, 1995.

....gives ffi n sp fi OPT ffiT , which is the length of time for EQUI s to complete the new fully parallelizable phase with work fi OPT ffiT . 11 Open Problems The performance of Equi partition has been studied extensively using simulation, experimental, and queuing theoretical approaches [3, 13, 16, 17, 24, 28, 2]. Our research constitutes a theoretical confirmation of these efforts. The main open problem is to close the gaps between the lower bounds on the competitive ratio known for general non clairvoyant schedulers and those known for the specific schedulers Equi partition and Balance. This gap is ....

C. McCann and J. Zahorjan. Scheduling memory constrained jobs on distributed memory parallel computers. In Proceedings of International Joint Conference on Measurement and Modeling of Computer Systems, ACM SIGMETRICS 95 and Performance 95, pages 208--219, 1995.

No context found.

McCann, C. and J. Zahorjan (1995). Scheduling Memory Constrained Jobs on Distributed Memory Parallel Computers. In ACM SIGMETRICS Conf. Measurement and Modeling of Computer Systems, pp. 208--219.

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