Adaptive parallelism on multiprocessors: Preliminary experience with Piranha on the CM-5

Abstract

Mechanisms for sharing multiprocessors among users are still in their infancy—typical approaches include simple space-sharing and inefficient, restricted forms of time-sharing. In this work we investigate a new alternative: adaptive parallelism [2]. Adaptively parallel programs can execute over a dynamically changing set of processors; many such codes can easily and dynamically share a multiprocessor by individually adapting to execute in separate groups of processors that may vary with time. Adaptive parallelism has been successfully used in Piranha, an execution model for Linda2 programs that turns idle networked workstations into a significant computing resource [4]. This work explores Piranha on the Connection Machine CM-5 multiprocessor. Our preliminary results suggest that adaptive parallelism can provide an attractive alternative to traditional methods for sharing multiprocessors among users.

This work is supported by Air Force Grant AFOSR-91-0098 and NASA Graduate Research Fellowship NGT-50858. CM-5 resources were provided by the Pittsburgh Supercomputing Center and the National Center for Supercomputing Applications.