Performance Gains from Leaving Idle Processors in Multiprocessor Systems.
Evgenia Smirni, Emilia Rosti, Giuseppe Serazzi, Lawrence W. Dowdy, Kenneth C. Sevcik · 1995
When a parallel job arrives at a multiprocessor and there are idle processors, it is a common practice to assign all available processors to the job. In this paper the performance advantages of not allocating all available processors in multiprocessor systems are investigated. The class of adaptive space-sharing policies is considered. Analytical models of simple adaptive policies are examined. Complex policies are presented and investigated by means of experiments on a 512 node Intel Paragon. For all experiments, real parallel applications are used as the system workload. Results are reported for single and multiclass cases. Sensitivity analysis with respect to the workload interarrival time distribution is presented. Performance gains derived from saving computational power (i.e., leaving idle processors while there are parallel jobs in the waiting queue) are identified as a function of the system size and workload characteristics. The more variable the workload characteristics are, ...