An evaluation of parallel job scheduling for ASCI Blue-Pacific
SC '99 Proceedings of the 1999 ACM/IEEE conference on Supercomputing
IEEE Software
On Job Scheduling for HPC-Clusters and the dynP Scheduler
HiPC '01 Proceedings of the 8th International Conference on High Performance Computing
A Gang-Scheduling System for ASCI Blue-Pacific
HPCN Europe '99 Proceedings of the 7th International Conference on High-Performance Computing and Networking
IPPS '99/SPDP '99 Proceedings of the 13th International Symposium on Parallel Processing and the 10th Symposium on Parallel and Distributed Processing
CCS Resource Management in Networked HPC Systems
HCW '98 Proceedings of the Seventh Heterogeneous Computing Workshop
On Advantages of Grid Computing for Parallel Job Scheduling
CCGRID '02 Proceedings of the 2nd IEEE/ACM International Symposium on Cluster Computing and the Grid
Utilization and Predictability in Scheduling the IBM SP2 with Backfilling
IPPS '98 Proceedings of the 12th. International Parallel Processing Symposium on International Parallel Processing Symposium
A Self-Tuning Job Scheduler Family with Dynamic Policy Switching
JSSPP '02 Revised Papers from the 8th International Workshop on Job Scheduling Strategies for Parallel Processing
Capacity estimation in HPC systems: simulation approach
ICDCIT'11 Proceedings of the 7th international conference on Distributed computing and internet technology
Enhancements to the decision process of the self-tuning dynp scheduler
JSSPP'04 Proceedings of the 10th international conference on Job Scheduling Strategies for Parallel Processing
| Hi-index | 0.00 |
In modern resource management systems for supercomputers and HPC-clusters the job-scheduler plays a major role in improving the performance and usability of the system. The performance of the used scheduling policies (e.g. FCFS, SJF, LJF) depends on the characteristics of the queued jobs. Hence we developed the dynP scheduler family. The basic idea was to change between different scheduling policies during runtime. The basic dynP scheduler uses the average estimated runtime of all queued jobs together with two input parameters to decide when a policy change may be benefical. A disadvantage is that the performance of the basic dynP scheduler strongly depends on the right setting of the two input parameters.Therefore we present the self-tuning dynP scheduler, which is totally independent from any parameter values. The basic concept is that the self-tuning dynP scheduler computes the full (virtual) schedule for each of the three policies in every scheduling step. Each computed schedule is rated by a criterion. Then the scheduler switches to that policy which generated the best schedule for the currently queued jobs. In this paper we are using simulations with trace based job sets to evaluate the performance of the scheduler. The achieved results are reasonably good compared to the parameterized dynP variant and the basic policies FCFS, SJF, and LJF.