A Parallel Workload Model and its Implications for Maui Scheduling Policies

We develop a workload model based on the logs of real workloads of the Linux cluster Atlas and a small IBM Blue Gene/L cluster at Lawrence Livermore National Laboratory (LLNL), and the 184-node IBM eServer pSeries 655/690 at the San Diego Supercomputer Center (SDSC). This model gives us insight into the performance of scheduling jobs on space-sharing parallel computers, provided by Maui. We find out that backfill queuing policies improve a lot of the system performance, and without reservation, the allocation policies do not have an obvious distance between each other.

[1]  Ramin Yahyapour,et al.  Parallel Computer Workload Modeling with Markov Chains , 2004, JSSPP.

[2]  Jie Chen,et al.  Parallel Job Scheduling with Overhead: A Benchmark Study , 2008, 2008 International Conference on Networking, Architecture, and Storage.

[3]  J. Banks,et al.  Discrete-Event System Simulation , 1995 .

[4]  Dror G. Feitelson,et al.  The workload on parallel supercomputers: modeling the characteristics of rigid jobs , 2003, J. Parallel Distributed Comput..

[5]  Mark S. Squillante,et al.  The impact of job arrival patterns on parallel scheduling , 1999, PERV.

[6]  Mark J. Clement,et al.  Core Algorithms of the Maui Scheduler , 2001, JSSPP.

[7]  Uwe Schwiegelshohn,et al.  Theory and Practice in Parallel Job Scheduling , 1997, JSSPP.

[8]  Mary K. Vernon,et al.  Evaluation of parallel job scheduling policies for production scientific workloads , 2001 .

[9]  Francine Berman,et al.  A model for moldable supercomputer jobs , 2001, Proceedings 15th International Parallel and Distributed Processing Symposium. IPDPS 2001.

[10]  Allen B. Downey,et al.  A parallel workload model and its implications for processor allocation , 1996, Proceedings. The Sixth IEEE International Symposium on High Performance Distributed Computing (Cat. No.97TB100183).