Harnessing the Effects of Process Variability to Mitigate Aging in Cloud Servers

The increasing number of cores within a single chip enables more powerful cloud servers to exploit request-level parallelism better. However, this also leads to unforeseen temperature issues, which may accelerate aging and cause soft-errors or even failures. Therefore, smartly managing temperature has become critical and unpredictable due to the inherent process variability - since the temperature will vary across the cores regardless if they operate at the same operating frequency. Based on that, we propose a framework to optimize cloud servers’ lifetime, which automatically distributes the workload among the cores and applies DVFS based on the applications’ behavior and systems status.

I. Introduction

The growing software-as-a-service demand has increased the pressure on warehouse infrastructures to support more robust cloud services, which involve applications from many domains, such as machine learning, biomedical, and video/audio processing. In such clouds, the workloads are highly heterogeneous and often result from requests from different clients. In such systems, the main challenge lies in providing the service with the lowest possible latency and wisely using the available resources while exploiting Request-Level Parallelism (RLP).