Performance Optimization of Many-Core Systems by Exploiting Task Migration and Dark Core Allocation
Shengyan Wen,Xiaohang Wang,Amit Kumar Singh,Yingtao Jiang,Mei Yang
DOI: https://doi.org/10.1109/tc.2020.3042663
IF: 3.183
2022-01-01
IEEE Transactions on Computers
Abstract:As an effective scheme often adopted for performance tuning in many-core processors, task migration provides an opportunity for "hot" tasks to be migrated to run on a "cool" core that has a lower temperature. When a task needs to migrate from one processor core to another, the migration can embark on numerous modes defined by the migration paths undertaken and/or the destinations of the migration. Selecting the right migration mode that a task shall follow has always been difficult, and it can be more challenging with the existence of dark cores that can be called back to service (reactivated), which ushers in additional task migration modes. Previous works have demonstrated that dark cores can be placed near the active cores to reduce power density so that the active cores can run at higher voltage/frequency levels for higher performance. However, the existing task migration schemes neither consider the impact of dark cores on each application's performance, nor exploit performance trade-off under different migration modes. Unlike the existing task migration schemes, in this article, a runtime task migration algorithm that simultaneously takes both migration modes and dark cores into consideration is proposed, and it essentially has two major steps. In the first step, for a specific migration mode that is tied to an application whose tasks need to be migrated, the number of dark cores is determined so that the overall performance is maximized. The second step is to find an appropriate core region and its location for each application to optimize the communication latency and computation performance; during this step, focus is placed on reducing the fragmentation of the free core regions resulting from the task migration. Experimental results have confirmed that our approach achieves over 50 percent reduction in total response time when compared to recently pr-posed thermal-aware runtime task migration approachess.
engineering, electrical & electronic,computer science, hardware & architecture