Defragmentation for Efficient Runtime Resource Management in NoC-Based Many-Core Systems

摘要

Efficient runtime resource allocation is critical to the overall performance and energy consumption of many-core systems. A region of free cores is allocated for each newly launched application. The cores are deallocated when the corresponding applications finish execution. The frequent allocations and deallocations of the cores might leave free cores scattered (not forming a contiguous region). This situation is referred to as fragmentation. Fragmentation could cause the inefficient mapping of the incoming applications, i.e., long communication distance between communicating cores. This further leads to poor performance and high energy consumption. In this paper, we propose a runtime defragmentation scheme that collects and reallocates the scattered cores in close proximity. We first define a fragmentation metric that is able to evaluate the scatteredness level of the free cores. Based on this, the proposed algorithm is executed to bring the scattered free cores together when the fragmentation metric is over a certain predefined threshold. In this way, the contiguous free core region is formed to facilitate the efficient mapping of the incoming applications. Moreover, the proposed algorithm also aims to minimize the negative impact on the performance of existing applications. Experimental results show that the proposed defragmentation scheme reduces the overall execution time and the energy consumption by 42% and 41%, respectively, when it is augmented to existing runtime mapping algorithms. Moreover, a negligible overhead, accounting for only less than 2.6% of the overall execution time, is required for the proposed defragmentation process. The proposed defragmentation scheme is an effective resource management enhancement to existing runtime mapping algorithms for many-core systems.