Postponing wearout failures in chip multiprocessors using thermal management and thread migration

This paper presents an improved method to postpone wearout failures and improves functional unit and entire system lifetime by considering two important wearout factors: temperature and functional unit usage. Our method provides a more fine grained approach as compared to prior methods by considering individual functional unit usage. Using this information, system behavior can be predicted and appropriate thread scheduling and migration decisions can be made. Our method incorporates temperature predictions based on recent historical temperatures and functional unit usages to rank threads and cores in a chip multiprocessor (CMP). Using these rankings, our method migrates threads among cores to reduce thermal hotspots. Simulation results on the ESESC simulator show that our method can improve the average system temperature and lifetime by approximately 4.33°C and 21.65%,respectively,in a tri-core CMP, and 6.4°C and 32% in a quad-core CMP.