PEACH: a model for performance and energy aware cooperative hybrid computing

Abstract

Accelerator-based heterogeneous systems become increasingly important to high performance computing because of their potentials to deliver high performance and energy efficiency. To fully realize this potential, parallel software must utilize both host processors and accelerators' computing power and power-aware capabilities. We develop PEACH, a model for Performance and Energy Aware Cooperative Hybrid computing. PEACH explores judicious workload distribution between hosts and accelerators and intelligent energy-aware scheduling for further performance and energy efficiency gains on heterogenous systems. With a few system- and application-dependent parameters, PEACH accurately captures the performance and energy impact of workload distribution and energy-aware scheduling.