Heterogeneity-Aware Peak Power Management for Accelerator-Based Systems

摘要

Power management has become one of the first-order considerations in high performance computing field. Many recent studies focus on optimizing the performance of a computer system within a given power budget. However, most existing solutions adopt fixed period control mechanism and are transparent to the running applications. Although the application-transparent control mechanism has relatively good portability, it exhibits low efficiency in accelerator-based heterogeneous parallel systems. In typical accelerator-based parallel systems, different processing units have largely different processing speeds and power consumption. Under a given power constraint, how to choose the processor to be slowed down and how to schedule a parallel task onto different processors for the maximum performance are different from those in homogeneous systems and have not been well studied. From the motivating example in this paper, we could find that in order to efficiently harness the heterogeneous parallel processing, one should not only perform dynamic voltage/frequency scaling (DVFS) to meet the power budget, but also tune the parallel task scheduling to adapt to the changes. In this paper, we propose a heterogeneity-aware peak power management, which extends existing application-transparent power controller with an application-aware power controller. Firstly, we theoretically analyze the conditions for the maximum performance given a power budget for heterogeneous systems. Based on this result, we provide a power-constrained parallel task partition algorithm, which coordinates parallel task partition and voltage scaling for heterogeneous processing units to achieve the optimal performance given a system power budget. Finally, we evaluate the proposed method on a typical CPU-GPU heterogeneous system, and validate the superiority of application-aware power controller over the existing method.