Automatic Self-Allocating Threads (ASAT) on the Convex Exemplar

Parallel processing systems have an advantage over traditional supercomputers in price/performance, but traditional supercomputers retain a signiicant advantage over parallel processing systems in the area of exibility. Traditional supercomputers can easily handle a mix of interactive, batch, scalar, vector, parallel, and large memory jobs simultaneously while maintaining high utilization. Often parallel processing systems do not eeectively support dynamic sharing of the resources of a system which results in low overall utilization given a mix of jobs. This paper describes the implementation of thread balancing software intended to allow the sharing of a large parallel processing system under a variety of load conditions. The system under consideration for this study is the Convex Ex-emplar scalable, parallel-processing system. While this paper is focused on the performance of this technique on the Convex Exemplar, it is a general technique which should prove to be useful on many shared memory parallel processors.