Power-Normalized Performance Optimization of Concurrent Many-Core Applications

Abstract

Modern operating systems, such as Linux, are capable of executing multiple parallel applications concurrently on many-core platforms. Different applications may have different characteristics with regard to how they exercise the computation and memory resources in these platforms. This paper aims to investigate the impact of such differences on the overall energy consumption and performance tradeoffs. To analyze these tradeoffs, three PARSEC benchmark applications are chosen with different characteristics - memory-intensive, CPU-intensive and a mixture of both. These applications are then concurrently executed in various combinations in experiments, which also help establish optimized run-time controls in terms of dynamic voltage/frequency scaling (DVFS) and thread-to-core allocations at run-time. Such controls are based on state-space models derived through linear regression using the feedback from hardware performance counters. Using the benchmark applications, we demonstrate the effectiveness of our proposed method, which shows up to 23% improvement in power normalized performance expressed as the ratio between instructions per second (IPS) and power consumption (Watt).