An optimal multi-processor allocation algorithm for high performance GPU accelerators

Abstract

An optimal algorithm for Multi-Processor Allocation in GPU system that reduce power consumption while maintain the application required performance is presented in this paper. Power consumption and heat dissipation have become critical issues in modern high performance computing systems due to the rising cost of electricity and the cooling infrastructure. The Multi-Processor Allocation (MPAlloc) algorithm will determine the appropriate number of Multi-Processor at runtime that can reduce power consumption, resources over-provisioning, and maintain performance simultaneously. It uses the memory BandWidth Utilization (BWU) metric to predict the MultiProcessor requirements of the application. The experimental results showed that a 14.2% of power saving could be achieved using the MPAlloc algorithm with less than 1% of performance degradation.