DASH: Dynamic Scheduling Algorithm for Single-ISA Heterogeneous Nano-scale Many-Cores

Abstract

The difference in the performance of identical cores due to the process variation and the different distance of cores from the shared last level cache (LLC) besides power limitation should be considered in scheduling algorithms to exploit the maximum performance of the Nano-scale many-core processors. This paper presents a dynamic heuristic scheduling algorithm, called DASH, to maximize performance under the mentioned challenges. In this regard, we estimate the execution time of each task of a job on a core as a relation of frequency and the communication cost of the core, as well as the type of the job and its tasks. According to this estimation model DASH selects some cores to maximize performance while exploiting DVFS to reduce the effect of power limitation. The time overhead of our algorithm is compatible with dynamic systems. We evaluate DASH by running random sequences of jobs from SPLASH parallel benchmark suite in Sniper and MACPat simulator for performance and power estimation. The results show that the throughput of DASH is 7.1% and 30.4% more than two similar algorithms.