Adjusting the Lengths of Time Slices when Scheduling PVM Jobs with High Memory Requirements

Abstract

Our research is focussed on keeping both local and parallel jobs together in a non-dedicated cluster and scheduling them efficiently. In such systems, the overflow of the physical memory into the virtual memory usually causes a severe performance penalty for distributed jobs. This impact can be reduced by means of giving more time slice length to parallel tasks in order better to exploit their memory reference locality. Thus, an algorithm is presented to adjust the length of the time slice dynamically to the necessity of the distributed and local tasks. It is implemented in a Linux cluster and evaluated with PVM jobs.

This work was supported by the MCyT under contract TIC 2001-2592 and partially supported by the Generalitat de Catalunya -Grup de Recerca Consolidat 2001SGR- 00218.