Remote AVX Overhead: Detection and Mitigation

Abstract

Due to power constraints, recent Intel CPUs reduce their frequency when executing AVX2 and AVX-512 instructions. Often, this frequency reduction affects other applications as well, which reduces overall performance and prevents contemporary operating systems from fairly distributing system resources. In our work, we show that these problems are fundamental problems of power-limited computing. We analyze the problems and show a method to quantify the underlying AVX overhead. Based on our analysis, we then describe a set of operating system techniques to improve performance and scheduler fairness. Our results show the importance of active management of hardware-controlled frequency scaling by the OS. Based on this observation, we sketch improved hardware-software interfaces which could further reduce AVX overhead and improve the efficacy of our approach.