Title

Adaptive space-time sharing with SCOJO.

Date of Award

2004

Degree Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

First Advisor

Sodan, A.,

Keywords

Computer Science.

Rights

CC BY-NC-ND 4.0

Abstract

Coscheduling is a technique used to improve the performance of parallel computer applications under time sharing, i.e., to provide better response times than standard time sharing or space sharing. Dynamic coscheduling and gang scheduling are two main forms of coscheduling. In SCOJO (Share-based Job Coscheduling), we have introduced our own original framework to employ loosely coordinated dynamic coscheduling and a dynamic directory service in support of scheduling cross-site jobs in grid scheduling. SCOJO guarantees effective CPU shares by taking coscheduling effects into consideration and supports both time and CPU share reservation for cross-site job. However, coscheduling leads to high memory pressure and still involves problems like fragmentation and context-switch overhead, especially when applying higher multiprogramming levels. As main part of this thesis, we employ gang scheduling as more directly suitable approach for combined space-time sharing and extend SCOJO for clusters to incorporate adaptive space sharing into gang scheduling. We focus on taking advantage of moldable and malleable characteristics of realistic job mixes to dynamically adapt to varying system workloads and flexibly reduce fragmentation. In addition, our adaptive scheduling approach applies standard job-scheduling techniques like a priority and aging system, backfilling or easy backfilling. We demonstrate by the results of a discrete-event simulation that this dynamic adaptive space-time sharing approach can deliver better response times and bounded relative response times even with a lower multiprogramming level than traditional gang scheduling.Dept. of Computer Science. Paper copy at Leddy Library: Theses & Major Papers - Basement, West Bldg. / Call Number: Thesis2004 .H825. Source: Masters Abstracts International, Volume: 43-01, page: 0237. Adviser: A. Sodan. Thesis (M.Sc.)--University of Windsor (Canada), 2004.