R. Agrawal, R.K. Mansharamani, M.K. Vernon, "Response time bounds for parallel processor allocation policies," Technical Report # 1152, Computer Science Dept., Univeristy of Wisconsin, Madison, WI, June 1993.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....completes or a new job arrives, whereas RTC policies cannot. Thus, the question of how many processors to allocate to a job is a particularly crucial issue in RTC scheduling. The optimal parallel processor scheduling policy is unknown except for a few very specific workload and system assumptions [AMV93,Se93]. Thus, many recent studies have compared specific policies to understand, for given workloads, which parallel scheduling policies perform better than others as well as which characteristics of scheduling policies lead to improved performance. For example, several studies have shown that among ....

....## =75; b 70,200,300,10000 ; r =0 4.6. Unification of Previous Results The experiments in sections 4.1 4.5 indicate that, for C v 1 application execution rate characteristics are not very useful for determining processor allocations in RTC policies. This is also true for exponential demands [AMV93]. In retrospect, the intuition behind this result is that policy performance degrades when jobs with long execution times are allocated too many processors. Since execution rate characteristics have not been shown or assumed to be correlated with demand, they are poor predictors of this behavior. ....

R. Agrawal, R. Mansharamani, M. Vernon. Response Time Bounds for Parallel Processor Allocation Policies. Technical Report 1152, Computer Sciences Dept., University of Wisconsin-Madison, June 1993.

....FCFS, EQuiallocation (EQ) policies, and Preemptive Smallest Available Parallelism First (PSAPF) perform almost the same [14, 18] ffl PSAPF is optimal for a workload with i.i.d. exponential task service times [13] and also for a workload with i.i.d. exponential job demands and linear speedups [1], given that the scheduler has no information about job processing requirements. ffl Under specific hyperexponential demands and specific parallelism distributions, PSAPF has been observed to have high performance under a workload with high correlation between demand and parallelism when ....

....policies have been shown to have high performance under various workloads [39, 14, 20, 8, 33, 21, 22, 18] The EQS policy is an idealized spatial equiallocation policy. Finally, 4 we examine the PSAPF policy that was proposed in [15] and shown to have high performance for specific workloads in [15, 13, 1]. Each of the policies is defined in the context of a global or central job queue. The four policies are defined in terms of the processing power that they allocate to jobs in the queue, and not in terms of the allocation of processors to individual tasks within a job. All four policies make use ....

[Article contains additional citation context not shown here]

R. Agrawal, R. Mansharamani, and M. Vernon. Response time bounds for parallel processor allocation policies. Technical Report # 1152, Computer Sciences Department, University of Wisconsin-Madison, June 1993.

....These assumptions are made for analytic tractability yet the solution techniques involve explicit enumeration of the state space, which yields no direct insight and grows exponentially in the number of processors. In this paper recent sample path analyses and interpolation approximation techniques [1, 16] are extended to analyze an idealization of the Spatial EQuiallocation policy (EQS) defined in Section 2) for a workload model that we believe is broadly applicable and uses only a few parameters to characterize the essential properties of parallel applications with respect to scheduling ....

....for REQS holds under more general workload assumptions, which include general job arrival times, job demands, available parallelisms, and execution rates, with arbitrary dependencies among these workload variables. The lower and upper bounds in this section are generalizations of the bounds in [1] for the EQS policy, and are obtained as corollaries of more general bounds, which show that the performance of EQS improves with increase in available parallelism. For example, for workloads with the same available parallelism for all jobs REQS decreases as available parallelism increases. In ....

[Article contains additional citation context not shown here]

R. Agrawal, R. Mansharamani, and M. Vernon. Response Time Bounds for Parallel Processor Allocation Policies. Technical Report # 1152, Computer Sciences Department, University of Wisconsin-Madison, June 1993.

No context found.

R. Agrawal, R.K. Mansharamani, M.K. Vernon, "Response time bounds for parallel processor allocation policies," Technical Report # 1152, Computer Science Dept., Univeristy of Wisconsin, Madison, WI, June 1993.

No context found.

R. Agrawal, R.K. Mansharamani, M.K. Vernon, "Response time bounds for parallel processor allocation policies," Technical Report # 1152, Computer Science Dept., Univeristy of Wisconsin, Madison, WI, June 1993.

No context found.

R. Agrawal, R.K. Mansharamani, M.K. Vernon, "Response time bounds for parallel processor allocation policies," Technical Report # 1152, Computer Science Dept., Univeristy of Wisconsin, Madison, WI, June 1993.

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