On the combined scheduling of malleable and rigid jobs

Abstract

The demand of the users of parallel systems for low response times contradicts the ambition of the system maintainers for a high utilization. A high utilization normally results in long waiting times for the users' jobs. To fullfil the concerns of both interest groups is a hard job to do. The usage of more flexible jobs models can be a way out of the dilemma. These models allow jobs to change their width at application start (moldable jobs) or even during execution (malleable jobs). We have analyzed the quality of schedules using job sets with moldable and malleable jobs and combinations of both. Tracefiles from supercomputer installations have been modified to contain varying fractions of moldable and malleable jobs. Using a special simulation environment for the more flexible job models the jobs have been scheduled virtually. The results show that both interest groups mentioned above can be pleased if these job models are used and the average response times become significantly better.