Improving Speedup and Response Times by Replicating Parallel Programs on a SNOW

Abstract

Idle computation cycles of a shared network of workstations are increasingly being used to run batch parallel programs. For one common paradigm, the batch program task running on an idle workstation is preempted when the owner reclaims the workstation. This owner interference has a considerable impact on the execution time of a batch program, especially in the case of large parallel programs. Replication of batch program tasks has been used to reduce the impact of owner interference. We show analytically that reclamation can significantly improve parallel program speedup. Perhaps surprisingly, replication can also improve efficiency for certain workloads. We present analysis to quantify the amount of speedup and efficiency improvement. Furthermore, we provide analysis to help determine whether extra available workstations should be used for increasing job parallelism or for task replication.

Scheduling Task Parallel Applications for Rapid Turnaround on Enterprise Desktop Grids

“…The authors in [23] use a probabilistic model similar to the one we described in Section 6 to analyze various replication issues. The platform model used in the study is similar to ours in that the resources are shared, task preemption is disallowed, and checkpointing is not supported.…”

A survey of task scheduling in Desktop Grid

“…В работе [60] доказано, что по критерию минимизации времени выполнения, репликация целесообразна только в «хвосте» вычисле-ний, то есть на том этапе, когда число заданий становится меньше числа вычислительных узлов (последний этап вычислений в рамках проекта). Оптимизация «хвоста» анализируется в работе [38].…”

Fault-aware grid scheduling using performance prediction by workload modeling

“…In their model, similar to [13,14], they assume that parallel tasks are only assigned to idle machines, which is not the case in grid environments, especially with existence of multiple scheduling systems (local scheduler in each site) and uniqueness of some resources. Although, there are few works in the area of performance modeling of desktop grids, [16][17][18][19], but in all of the above-mentioned works, there is no effort for impatient applications, whom should be serviced right away, and consideration of resources' fault in the performance modeling. In the area of scheduling, some researchers translated the job scheduling problem into linear programming or K-out-of-N rule formats, and mapped it onto an appropriate neural network structure by which they came up with a reasonable solution [20][21][22].…”