HARS, a Heterogeneity-Aware Runtime System for Self-Adaptive Multithreaded Applications

Abstract

We are in the age of rapid changes. In particular, the computer is not just a device for carrying out complex computations. Various users can use these devices in various ways. Someone could just use a computer for document writing, while others can use one for video encoding or playing computer games. Each process has different computational demands. If we use a supercomputer for writing documents only, it is truly a wasteful mistake. It is like using a sledgehammer to crack a nut. Each process has own computational demands. If we can give the proper computation power to each process, this would be more sufficient than the before. In this manner, heterogeneous multi-processing (HMP) arose. HMP is a promising technique that can support both high and low demand tasks efficiently. This topic has been investigated in some prior works but an efficient system software to support HMP with self-adaptive computing has been little researched, especially on multithreaded applications. Therefore, we propose HARS, a heterogeneity-aware runtime system for self-adaptive multithreaded applications. HARS monitors application-level performance data and dynamically controls the system state to achieve the performance target with efficient power consumption. As an extended version of HARS, we also propose MP-HARS, which that supports multiple applications. Through our evaluation, we can see that HARS and MP-HARS achieve higher efficiency then the baseline version and HARS is comparable to the static optimal version.