Processor allocation using user directives in mesh-connected multicomputer systems

Abstract

Contemporary processor allocation schemes for multicomputers suffer from a fragmentation problem which causes underutilization of the processing nodes. The RSR and ANCA schemes, based on the concepts of size-reduction and non-contiguous allocation, show considerable performance improvement. However the penalties associated with these schemes limit their usage in environments with memory-bounded or communication-intensive jobs. We propose to use simple directives provided by the users to apply the appropriate allocation scheme and thus maximize the performance benefit. We present such a hybrid processor allocation scheme which combines the previously proposed conventional allocation, RSR and ANCA schemes. The hybrid allocation scheme is evaluated via extensive simulation. It outperforms the conventional allocation scheme and can be implemented easily while maximizing the resource utilizations.