Bandwidth-Aware Resource Optimization for SMT Processors

Abstract

In this paper we have studied the memory bandwidth and throughput behaviour of various SPEC CPU2006 workloads in Single Threaded (ST) and Simultaneous Multi-threaded (SMT) environment with varying number of workload copies on IBM POWER7 processor. Our study reveals that for a number of workloads, the bandwidth gets saturated beyond a certain number of copies and the performance gain achieved by increasing the number of copies is very minimal after the bandwidth saturation point. The results also show that for such workloads, after bandwidth saturation, the similar performance as achieved with ST mode, can be achieved by moving the copies to SMT environment with reduced number of cores. This enables us to free-up cores which can be used to intelligently co-schedule non-memory intensive applications. This is of particular value in the HPC cloud environments and can help increase overall throughput of a cloud data-center or cluster. In a non-cloud or traditional environment, this approach has applicability in bringing out workload consolidation and better performance-per-watt. We have also demonstrated by actual hardware implementation of how the freed-up cores by moving to SMT can used effectively for doing other useful computations.