X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In SODA: ACMSIAM Symposium on Discrete Algorithms, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....the adversary can choose, increases the knowledge of the non clairvoyant scheduler, or increases the power of the scheduler in some way. By covering many of these issues, we bring many of these paths of research together. 1. 3 Advantages Given to the Non Clairvoyant Scheduler One line of research [25, 5, 7], advantages the non clairvoyant scheduler by considering only batch jobs, i.e. all the jobs are released at time zero. However, in most practical settings, this not the case. Hence, we allow the jobs to be released at arbitrary times. Some results consider only fully parallelizable jobs (or ....

....job was released and whether or not it has completed. The schedulers in some results are computationally intensive. Finding the optimal may be NPcomplete. Even when polynomial, the algorithm may (e.g. involving finding a perfect matching) not be practical in a real time situation. Other results [18, 5, 9, 7], like ours, consider only computationally simple algorithms. Preemptive scheduling allows the number of processors allocated to a job to be changed after the job starts its execution. This helps adapt to the uncertain and changing nature of jobs and workloads. Unfortunately, preemption may incur ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Seventh Annual ACM--SIAM Symposium on Discrete Algorithms, pages 159--167, Atlanta, Georgia, January 1996.

....the adversary can choose, increases the knowledge of the non clairvoyant scheduler, or increases the power of the scheduler in some way. By covering many of these issues, we bring many of these paths of research together. 1. 3 Advantages Given to the Non Clairvoyant Scheduler One line of research [25, 5, 7], advantages the non clairvoyant scheduler by considering only batch jobs, i.e. all the jobs are released at time zero. However, in most practical settings, this not the case. Hence, we allow the jobs to be released at arbitrary times. Some results consider only fully parallelizable jobs (or ....

....job was released and whether or not it has completed. The schedulers in some results are computationally intensive. Finding the optimal may be NPcomplete. Even when polynomial, the algorithm may (e.g. involving finding a perfect matching) not be practical in a real time situation. Other results [18, 5, 9, 7], like ours, consider only computationally simple algorithms. Preemptive scheduling allows the number of processors allocated to a job to be changed after the job starts its execution. This helps adapt to the uncertain and changing nature of jobs and workloads. Unfortunately, preemption may incur ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Seventh Annual ACM--SIAM Symposium on Discrete Algorithms, pages 159--167, Atlanta, Georgia, January 1996.

....not require a specific number of processors, but can use different numbers. However, once a partition for a job has been selected its size cannot change anymore. Finally, in dynamic partitioning the size of a partition may change at run time. This model has, for instance, been used by Deng et al. [11]. 2. Job Flexibility As already mentioned advanced partitioning methods must not only be supported by the multiprocessor system but by the application as well. Therefore, Feitelson and Rudolph [25] characterize applications as follows: Rigid jobs. The number of processors assigned to a job is ....

....Gang scheduling may be implemented with or without migration. While many theoretical scheduling studies only use a model without preemption, more recently preemption has also been taken into account. Schwiegelshohn [71] uses a gang scheduling model without migration. The work of Deng et al. [11] is based upon migratable preemption. In a real system the preemption of a job requires that all the job s threads be stopped in a consistent state (i.e. without any messages being lost) and the full state of each thread must be preserved. The memory contents associated with the job may be ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu, "Preemptive scheduling of parallel jobs on multiprocessors". In 7th SIAM Symp. Discrete Algorithms, pp. 159--167, Jan 1996.

....for Parallel Job Scheduling A DISSERTATION submitted in partial satisfaction of the requirements for the degree of DOCTOR OF PHILOSOPHY in Information and Computer Science by Raghu Subramanian Thesis Committee: Professor Isaac D. Scherson, Chair Professor Sandra Irani Professor Alexandru Nicolau 1995 Copyright c fl1995 by Raghu Subramanian All rights reserved The dissertation of Raghu Subramanian is approved and is acceptable in quality and form for publication on microfilm: Committee Chair University of California, Irvine 1995 ii Contents List of Figures : ....

....A DISSERTATION submitted in partial satisfaction of the requirements for the degree of DOCTOR OF PHILOSOPHY in Information and Computer Science by Raghu Subramanian Thesis Committee: Professor Isaac D. Scherson, Chair Professor Sandra Irani Professor Alexandru Nicolau 1995 Copyright c fl1995 by Raghu Subramanian All rights reserved The dissertation of Raghu Subramanian is approved and is acceptable in quality and form for publication on microfilm: Committee Chair University of California, Irvine 1995 ii Contents List of Figures : ....

[Article contains additional citation context not shown here]

X. Deng et al. Preemptive scheduling of parallel jobs in multiprocessors. Technical Report CS-95-04, York University, North York, Ontario, Canada, July 1995.

.... [3] Similarly, completion time scheduling algorithms with a constant approximation factor have been provided with respect to o line scheduling [8] randomized scheduling [9] preemptive scheduling [10] scheduling of moldable (malleable) jobs [11] and scheduling with unknown job execution times [12]. In general a parallel job schedule S determines for each job i its starting time and the subset of nodes M i assigned to this job. The starting time of job i must be greater than its submission time or release date r i and node subsets of two jobs executing concurrently must be disjoint. The ....

....with an approximation factor of 8.53. Chakrabarti et al. 9] derived a similar bound of 8.67 for a randomized method and an on line problem (unknown submission times) Schwiegelshohn [10] was able to obtain an improved bound of 2.37 for the o line problem by using preemption. Finally, Deng et al. [12] proved an approximation factor of 2 for their preemptive on line model (unknown job execution times) with unit weights. At this point we would like to address the di erence between weighted completion and weighted ow time ( completion time submission time) Obviously, the user is only ....

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Proceedings of the 7 th SIAM Symposium on Discrete Algorithms, pages 159-167, January 1996.

....may result in schedules which deviate significantly from the optimum. Other results include Chakrabarti et al. 1] who addressed for instance the scheduling of malleable jobs similar to [14] and proved an expected performance within 8.67 of optimal for a randomized on line algorithm. Deng et al. [2] also discussed preemptive response time scheduling for malleable jobs with unique weights and variable resource requirements. But their model is related to a multithreaded environment and differs significantly from ours as they allow unknown execution times while neither a preemption penalty nor ....

....Further, some special cases are considered in Section 7. Finally, our results are compared to the worst case bounds for non preemptive SMART schedules. 2 The Model A system of M independent parallel jobs must be executed on a multiprocessor. Each job has only a single phase of parallelism [2]. It is described by its invariant resource requirement r i 2 f1; Rg (also called fixed allotment in [14] its execution time h i 1, and a user set priority weight w i 0. Note, that without restriction of generality the minimal job execution time has been normalized to 1. Our ....

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Proceedings of the 7 th SIAM Symposium on Discrete Algorithms, January 1996.

....an approximation factor of 8.53. Chakrabarti et al. 1] derived a similar bound of 8.67 for a randomized method and an on line problem (unknown submission times) Schwiegelshohn [10] was able to obtain an improved bound of 2.37 for the off line problem by using preemption. Finally, Deng et al. [2] proved an approximation factor of 2 for their preemptive on line model (unknown job execution times) with unit weights. At this point we would like to address the difference between weighted completion and weighted flow time ( completion time Gamma submission time) Obviously, the user is only ....

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Proceedings of the 7 th SIAM Symposium on Discrete Algorithms, pages 159--167, January 1996.

....not require a specific number of processors, but can use different numbers. However, once a partition for a job has been selected its size cannot change anymore. Finally, in dynamic partitioning the size of a partition may change at run time. This model has, for instance, been used by Deng et al. [12]. 2. Job Flexibility As already mentioned advanced partitioning methods must not only be supported by the multiprocessor system but by the application as well. Therefore, Feitelson and Rudolph [26] characterize applications as follows: low User activity Time 8am 12pm 2pm 4pm 6pm 8pm 10pm 12am ....

....Gang scheduling may be implemented with or without migration. While many theoretical scheduling studies only use a model without preemption, more recently preemption has also been taken into account. Schwiegelshohn [72] uses a gang scheduling model without migration. The work of Deng et al. [12] is based upon migratable preemption. In a real system the preemption of a job requires that all the job s threads be stopped in a consistent state (i.e. without any messages being lost) and the full state of each thread must be preserved. The memory contents associated with the job may be ....

X. Deng, N. Gu, T. Brecht, and K. Lu, "Preemp- tive scheduling of parallel jobs on multiprocessors ". In 7th SIAM Symp. Discrete Algorithms, Jan 1996.

....(n) 63] Interestingly, the results for batch style scheduling can be generalized to parallel jobs. Here we consider only total completion time (note that waiting time is not well defined for malleable parallel jobs) For ideally malleable jobs there exists a 2 competitive deterministic algorithm [25], matching the performance for sequential jobs (consequently, randomization cannot help in this case) A wide range of types of non ideally malleable jobs together with various restrictions on the number of preemptions is studied in [29] It is even possible to obtain algorithm for non ideally ....

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Proc. of the 7th Ann. ACM-SIAM Symp. on Discrete Algorithms, pages 159--167. ACM-SIAM, 1996.

....different speeds. Many of the new results consider scheduling of jobs with unknown running times without dependencies with the objective to minimize the total (or average) completion time, rather than makespan; good bounds have been proved for sequential jobs [16] ideally malleable parallel jobs [5], and even for a rather general class of nonideally malleable jobs [4] Other papers study also the scenario in which the jobs arrive one by one, or arrive over time but with known running times; see the survey [20] for the references. 2 Definitions 2.1 Parallel machines and parallel job systems ....

Xiaotie Deng, Nian Gu, Tim Brecht, and KaiCheng Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Proc. of the 7th Ann. ACM-SIAM Symp. on Discrete Algorithms, pages 159--167. ACM-SIAM, 1996.

....algorithms for minsum open shop scheduling and minsum job shop scheduling with a fixed number of machines. Previous work on the WACT metric gave either non clairvoyant, preemptive solutions for sequential jobs or jobs that used a perfectly malleable processor resource, with job precedence [41]; or clairvoyant, non preemptive solutions for independent jobs with only non malleable resource, if any [92, 112] Apart from algorithm design, we believe it is important to point out some differences between existing scheduling literature and features needed by schedulers in parallel computing ....

....permitting out of order execution. 6. 2 Multiprocessor scheduling Job scheduling in multiprocessors has been extensively researched in both Operations Research and Computer Science [71] Two decades later there is significant interest in the Computer Science community on scheduling parallel jobs [111, 41]. For independent jobs that need a fixed given number of processors, various rectangle packing algorithms naturally model the problem: every job can be regarded as a rectangle of width equal to the number of processors, and of height representing the running time [14, 13] It is more realistic to ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Symposium on Discrete Algorithms (SODA). ACM-SIAM, Jan. 1996.

....the adversary can choose, increases the knowledge of the non clairvoyant scheduler, or increases the power of the scheduler in some way. By covering many of these issues, we bring many of these paths of research together. 1. 3 Advantages Given to the Non Clairvoyant Scheduler One line of research [25, 5, 7], advantages the non clairvoyant scheduler by considering only batch jobs, i.e. all the jobs are released at time zero. However, in most practical settings, this not the case. Hence, we allow the jobs to be released at arbitrary times. Some results consider only fully parallelizable jobs (or ....

....job was released and whether or not it has completed. The schedulers in some results are computationally intensive. Finding the optimal may be NPcomplete. Even when polynomial, the algorithm may (e.g. involving finding a perfect matching) not be practical in a real time situation. Other results [18, 5, 9, 7], like ours, consider only computationally simple algorithms. Preemptive scheduling allows the number of processors allocated to a job to be changed after the job starts its execution. This helps adapt to the uncertain and changing nature of jobs and workloads. Unfortunately, preemption may incur ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Seventh Annual ACM--SIAM Symposium on Discrete Algorithms, pages 159--167, Atlanta, Georgia, January 1996.

....the adversary can choose, increases the knowledge of the non clairvoyant scheduler, or increases the power of the scheduler in some way. By covering many of these issues, we bring many of these paths of research together. 1. 3 Advantages Given to the Non Clairvoyant Scheduler One line of research [25, 5, 7], advantages the non clairvoyant scheduler by considering only batch jobs, i.e. all the jobs are released at time zero. However, in most practical settings, this not the case. Hence, we allow the jobs to be released at arbitrary times. Some results consider only fully parallelizable jobs (or ....

....job was released and whether or not it has completed. The schedulers in some results are computationally intensive. Finding the optimal may be NPcomplete. Even when polynomial, the algorithm may (e.g. involving finding a perfect matching) not be practical in a real time situation. Other results [18, 5, 9, 7], like ours, consider only computationally simple algorithms. Preemptive scheduling allows the number of processors allocated to a job to be changed after the job starts its execution. This helps adapt to the uncertain and changing nature of jobs and workloads. Unfortunately, preemption may incur ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Seventh Annual ACM--SIAM Symposium on Discrete Algorithms, pages 159--167, Atlanta, Georgia, January 1996.

.... jobs remaining in the system a quantum of processor time proportional to how long they have been in the system [19] In multiprocessor systems a similar preemptive algorithm, dynamic Equi partition (DEQ) can be used to achieve similar performance when preemption costs are not prohibitively large [3, 4]. However, overheads incurred due to preemptive scheduling algorithms may be much larger in parallel and distributed systems, and especially in the networks of workstations model. When the overhead is prohibitive, then results from theoretical studies on non preemptive execution of parallel jobs ....

....when the jobs are from a fairly large class. The number of preemptions in Equi partition can be further reduced to log k n with an extra constant multiplicative factor of k loss in performance. This result is perhaps most interesting when compared with the existing bound (4 times optimal) [4] for the dynamic Equi partition algorithm (DEQ) Our new bound for Equi partition is tighter than the previous bound for DEQ, even though Equi partition uses significantly fewer preemptions and does not use any job execution characteristics, whereas DEQ does. That is, for the job class for which ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. SIAM J. Comput. 30(1): 145-160 (2000)

.... jobs remaining in the system a quantum of processor time proportional to how long they have been in the system [19] In multiprocessor systems a similar preemptive algorithm, dynamic Equi partition (DEQ) can be used to achieve similar performance when preemption costs are not prohibitively large [3, 4]. However, overheads incurred due to preemptive scheduling algorithms may be much larger in parallel and distributed systems, and especially in the networks of workstations model. When the overhead is prohibitive, then results from theoretical studies on non preemptive execution of parallel jobs ....

....when the jobs are from a fairly large class. The number of preemptions in Equi partition can be further reduced to log k n with an extra constant multiplicative factor of k loss in performance. This result is perhaps most interesting when compared with the existing bound (4 times optimal) [4] for the dynamic Equi partition algorithm (DEQ) Our new bound for Equi partition is tighter than the previous bound for DEQ, even though Equi partition uses signi cantly fewer preemptions and does not use any job execution characteristics, whereas DEQ does. That is, for the job class for which ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. SIAM J. Comput. 30(1): 145-160 (2000)

.... remaining in the system a quantum of processor time proportional to how long they have been in the system [15] In multiprocessor systems a similar preemptive algorithm, dynamic Equi partition (DEQ) can be used when preemption costs are not prohibitively large 1 , to achieve similar performance [2, 3]. However, overheads incurred due to preemptive scheduling algorithms may be much larger in parallel and distributed systems, and especially in the networks of workstations model. When the overhead is prohibitive, then results from theoretical studies on non preemptive execution of parallel jobs ....

....when the jobs are from a fairly large class. The number of preemptions in Equi partition can be further reduced to log 2 n with an extra constant multiplicative factor of two loss in performance. This result is perhaps most interesting when compared with the existing bound (4 times optimal) [3] for the dynamic Equi partition algorithm (DEQ) Our new bound for Equi partition is tighter than the previous bound for DEQ, even though Equi partition uses significantly fewer preemptions and does not use any job execution characteristics, whereas DEQ does. A possible interpretation of this ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Seventh Annual ACM--SIAM Symposium on Discrete Algorithms, pages 159--167, Atlanta, Georgia, January 1996.

.... jobs remaining in the system a quantum of processor time proportional to how long they have been in the system [13] In multiprocessor systems a similar preemptive algorithm, dynamic Equi partition (DEQ) can be used to achieve similar performance when preemption costs are not prohibitively large [3, 4]. However, overheads incurred due to preemptive scheduling algorithms may be much larger in parallel and distributed systems, and especially in the networks of workstations model. When the overhead is prohibitive, then results from theoretical studies on non preemptive execution of parallel jobs ....

....when the jobs are from a fairly large class. The number of preemptions in Equi partition can be further reduced to log k n with an extra constant multiplicative factor of k loss in performance. This result is perhaps most interesting when compared with the existing bound (4 times optimal) [4] for the dynamic Equi partition algorithm (DEQ) Our new bound for Equi partition is tighter than the previous bound for DEQ, even though Equi partition uses significantly fewer preemptions and does not use any job execution characteristics, whereas DEQ does. That is, for the job class for which ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In Seventh Annual ACM--SIAM Symposium on Discrete Algorithms, pages 159--167, Atlanta, Georgia, January 1996.

.... 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 memory access for multiprocessor systems: distributed memory and shared memory. 1 The approach of Equi partition is first introduced to parallel scheduling by Tucker and Gupta as a process control policy [27] and ....

.... 2 Gamma 2 n 1 times optimal and no policy can guarantee a better competitive ratio [20] For shared memory models, Deng et al. were able to show that DEQ achieves a solution within 2 Gamma 2 n 1 times the optimal average completion time for parallelism profile jobs with only one phase [5]. Surprisingly this is exactly the same competitive ratio as uniprocessor systems. They also showed that DEQ can achieve 4 Gamma 4 n 1 times the optimal average completion time for parallelism profile jobs with multiple phases. For distributed memory multiprocessor models, there is no ....

[Article contains additional citation context not shown here]

X. Deng, N. Gu, T. Brecht, and K. Lu "Preemptive Scheduling of Parallel Jobs on Multiprocessors," to be presented on The Seventh Annual ACM-- SIAM Symposium on Discrete Algorithms, Atlanta, January, 1996.

No context found.

X. Deng, N. Gu, T. Brecht, and K. Lu. Preemptive scheduling of parallel jobs on multiprocessors. In SODA: ACMSIAM Symposium on Discrete Algorithms, 1996.

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