Machine Learning based Electromigration-aware Scheduler for Multi-core Processors

Abstract

The rising performance demands in modern technology devices see the need to pack more functionality per area and are made possible with the advent of technology scaling. The extremely down-scaled, high-density processors used in such technology devices functioning at high frequencies and greater temperatures expedite various aging effects which impact the reliable lifetime of computing systems. Electromigration is considered to be an important intrinsic aging effect that reduces the useful lifetime of modern microprocessors. The objective of this work is to use machine learning methods to develop an electromigration-aware scheduler for assigning workloads to cores based on reliability and performance requirements. Aging estimation of the processor cores is performed based on the proposed computationally efficient and accurate regressionbased thermal prediction models. According to experimental findings, the suggested technique can significantly extend the lifetime of multi-core architectures while allowing performance to degrade gracefully. The maximum error in the prediction of the lifetime of the cores using the proposed methodology is estimated to be 2.85%.