Parsons EW, Sevcik KC. Implementing multiprocessor scheduling disciplines. Proceedings of the IPPS Workshop on Job Scheduling Strategies for Parallel Processing (JSSPP) (Lecture Notes in Computer Science, vol. 1291). Springer: Berlin, 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....[9] Thus timesharing between jobs only makes sense if all the jobs being timeshared fit within the memory of the host, which is unlikely. Also, many operating systems that enable timesharing for single processor jobs do not facilitate preemption among several processors in a coordinated fashion [21]. done based on the IP addresses chosen by the Domain Name Server. The Round Robin policy is simple, but it neither maximizes utilization of the hosts, nor minimizes response times. In the case where the size of jobs are not known and we require immediate dispatch of jobs to hosts, then the ....

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. In Proceedings of IPPS/SPDP '97 Workshop. Lecture Notes in Computer Science, vol. 1459, pages 166--182, April 1997.

....Thus timesharing between jobs only makes sense if all the jobs being timeshared fit within the memory of the host, which is very unlikely. Also, many operating systems that enable timesharing for single processor jobs, do not facilitate preemption among several processors in a coordinated fashion [21]. Examples of distributed server systems that fit the above description are given in Table 1. Lastly, we assume that no a priori information is known about the job at the time when the job arrives. In particular, the processing requirement of the job is not l nown. We will use the terms processing ....

Eric W. Parsons and Kenneth C. Sevcik. Implementing multiprocessor scheduling disciplines. In Proceedings of IPPS'/S'PDP '97 Workshop. Lecture Notes m Computer 3'ciece, vol. 1359, pages 166 182, April 1997.

....and where jobs are executed. Obviously, there has to be close cooperation between the scheduler and the underlying infrastructure. During the past, researchers often worked independently on these two layers. As a consequence, there is now a large gap between research results achieved in both areas [19,32]. The best known scheduling algorithms, were studied using analytical models that are not directly compatible with the infrastructure implemented in the various metacomputing projects. It is therefore the purpose of this work to show a possible way of bridging this gap and to qualify the ....

....those discussed in [29] Often, these strategies are difficult to be implemented but guarantee a good worst case behavior. On the other hand, there are several fairly simple heuristics that have been tested in simulation environments or sometimes even as part of real world scheduling systems [32,4,10]. These approaches do not have such a good worst case behavior, but the results for typical problems are often quite good [29] We decided to concentrate our examinations on those strategies that are not too complex for implementation and for which good performance has been demonstrated with ....

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. Lecture Notes in Computer Science, 1291:166ff, 1997.

....The time when a job will finish is often quite predictable. A popular scheduling algorithm of this type is EASY [SCZL96] which guarantees that jobs will never be delayed by jobs submitted after them. However, research has demonstrated the desirability of more adaptable disciplines and job models [PS97] 2.1.4 Space Partitioning Algorithms that divide the available processors among the competing jobs without explicitly time slicing processors among jobs are known as space sharing or space partitioning. Chapter 2. Background 20 They can be classified similarly to the type of job model that ....

....is dependent on the amount and type of messages received and sent by a process. Processes that are communicating often will increase their spin time. They give results showing this mechanism to be competitive with explicit coscheduling; a maximum slowdown of 30 is observed. Parsons and Sevcik [PS97] have done work implementing a range of scheduling disciplines on a 16 node NOW. The scheduler is implemented on top of the commercial load balancing system LSF. They show that despite high preemption and reconfiguration overheads (over a minute for preemption) adaptive and preemptive ....

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. In Job Scheduling Strategies for Parallel Processing -- IPPS '97 Workshop, volume 1291 of Lecture Notes in Computer Science, New York, NY, USA, 1997. Springer-Verlag Inc.

....discussed. Here, a job that fits idle processors means a job that requests processors whose number is not exceeding the number of idle processors. In Fit Processors First Served (FPFS) a job scheduler searches jobs in the job queue and dispatches a job that fits idle processors to processors [14 16] 1 . Backfilling is a similar scheme to FPFS. In Backfilling, when the job at the head of the job queue waits for being dispatched because there are not enough idle processors for it, a job scheduler dispatches other jobs that fit idle processors without affecting the start time of the job at ....

.... for it, a job scheduler dispatches other jobs that fit idle processors without affecting the start time of the job at the head of the job queue [13, 17, 18] Several job scheduling schemes described in the previous paragraph has been evaluated either by simulation or experiments on real machines [13, 16 18]. However, none of work has compared performance of all these job scheduling schemes and has evaluated their performance by an analytical method. This paper evaluates the performance of job scheduling schemes for pure space sharing among rigid jobs. First, this paper evaluates and compares the ....

[Article contains additional citation context not shown here]

E. W. Parsons and K. C. Sevcik. Implementing Multiprocessor Scheduling Disciplines. In Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science 1291, pages 166--192. Springer-Verlag, 1997.

....are not well understood, and often rely on adhoc parameters. The Web pages are very vague on this issue and are often contradicted by users of these systems [6] The schedulers used are Load Leveler, LSF, PBS, or NQS. These schedulers typically only support run to completion (no preemption) [16]. In several distributed servers we looked at, the users submit an upper bound on the processing requirement of their job. In some systems task assignment is determined by the user. The jobs at each machine 1 For example, Processor Sharing (which requires infinitely many preemptions) is ....

Eric W. Parsons and Kenneth C. Sevcik. Implementing multiprocessor scheduling disciplines. In Proceedings of IPPS/SPDP '97 Workshop. Lecture Notes in Computer Science, vol. 1459, pages 166--182, April 1997.

....Thus timesharing between jobs only makes sense if all the jobs being timeshared fit within the memory of the host, which is very unlikely. Also, many operating systems that enable timesharing for single processor jobs, do not facilitate preemption among several processors in a coordinated fashion [21]. Examples of distributed server systems that fit the above description are given in Table 1. Lastly, we assume that no a priori information is known about the job at the time when the job arrives. In particular, the processing requirement of the job is not known. We will use the terms processing ....

Eric W. Parsons and Kenneth C. Sevcik. Implementing multiprocessor scheduling disciplines. In Proceedings of IPPS/SPDP '97 Workshop. Lecture Notes in Computer Science, vol. 1459, pages 166--182, April 1997.

....job are considered to be a single gang. Thus the number of threads in the job conveys the PE requirements of the job. The simple version of gang scheduling, where threads are mapped to PEs and do not move, is currently the most popular one. However, more flexible versions have also been proposed [449]. One of these is migratable preemption, whereby threads are preempted on one set of PEs and resume on another (typically disjoint) set. This may be used to reduce fragmentation. A more drastic version is malleable preemption, in which the job is resumed on a set of PEs with a different size from ....

....in an interactive time scale. An interesting observation relating to gang scheduling via swapping is that it is related to checkpointing [365] As such, it is possible to restart the job on a different set of PEs than the one used originally. Therefore swapping leads to migratable preemption [449], and can help in the reduction of fragmentation [346, 516] While simultaneous swapping is a type of gang scheduling, it need not be the only one. It is possible to have both fine grain gang scheduling among the memory resident jobs, and coarse grain gang scheduling by means of swapping. A ....

[Article contains additional citation context not shown here]

E. W. Parsons and K. C. Sevcik, "Implementing multiprocessor scheduling disciplines". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 166--192, Springer Verlag, 1997. Lecture Notes in Computer Science Vol. 1291.

....in an interactive time scale. An interesting observation relating to gang scheduling via swapping is that it is related to checkpointing [220] As such, it is possible to restart the job on a different set of PEs than the one used originally. Therefore swapping leads to migratable preemption [275], and can help in the reduction of fragmentation [209, 311] While simultaneous swapping is a type of gang scheduling, it need not be the only one. It is possible to have both fine grain gang scheduling among the memory resident jobs, and coarse grain gang scheduling by means of swapping. A ....

E. W. Parsons and K. C. Sevcik, "Implementing multiprocessor scheduling disciplines". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 166--192, Springer Verlag, 1997. Lecture Notes in Computer Science Vol. 1291.

....with the idea of using available service time knowledge to favor the jobs that are expected to complete most promptly. If better knowledge of job service times than queue identities is available, then it is best to try to activate the jobs in order of increasing expected remaining service time [63]. If the service times are known to be highly variable, but the service times of individual jobs cannot be predicted in advance, then the discipline that executes the job with least acquired service first is best because it emulates the behavior of least expected remaining work first. ....

E. W. Parsons and K. C. Sevcik, "Implementing multiprocessor scheduling disciplines ". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), Springer Verlag, 1997. Lecture Notes in Computer Science (this volume).

....with the idea of using available service time knowledge to favor the jobs that are expected to complete most promptly. If better knowledge of job service times than queue identities is available, then it is best to try to activate the jobs in order of increasing expected remaining service time [64]. If the service times are known to be highly variable, but the service times of individual jobs cannot be predicted in advance, then the discipline that executes the job with least acquired service first is best because it emulates the behavior of least expected remaining work first. ....

E. W. Parsons and K. C. Sevcik, "Implement- ing multiprocessor scheduling disciplines". In IPPS'97 Workshop Job Scheduling Strategies for Parallel Processing.

No context found.

Parsons EW, Sevcik KC. Implementing multiprocessor scheduling disciplines. Proceedings of the IPPS Workshop on Job Scheduling Strategies for Parallel Processing (JSSPP) (Lecture Notes in Computer Science, vol. 1291). Springer: Berlin, 1997.

No context found.

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. In Proceedings of IPPS/SPDP '97 Workshop. Lecture Notes in Computer Science, vol. 1459, pages 166--182, April 1997.

No context found.

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. In Proceedings of IPPS/SPDP '97 Workshop. Lecture Notes in Computer Science, vol. 1459, pages 166--182, April 1997.

No context found.

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. In D. G. Feitelson and L. Rudolph, editors, Job Scheduling Strategies for Parallel Processing, volume 1291 of Lecture Notes in Computer Science, pages 166--192, Berlin Heidelberg New York, 1997. Springer-Verlag.

No context found.

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. In Proceedings of IPPS/SPDP '97 Workshop. Lecture Notes in Computer Science, vol. 1459, pages 166--182, April 1997.

No context found.

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. In D. G. Feitelson and L. Rudolph, editors, Job Scheduling Strategies for Parallel Processing, pages 166--192. Springer Verlag, 1997. Lect. Notes Comput. Sci. vol. 1291.

No context found.

E. W. Parsons and K. C. Sevcik. Implementing multiprocessor scheduling disciplines. Lecture Notes in Computer Science, 1291:166#, 1997.

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