Understanding and Using Heterogeneity for High Performance, Energy Efficient Computing: Special Session Extended Abstract

This paper identifies workload and platform heterogeneity as an important feature that needs to be modeled and exploited for optimizing for performance and energy efficiency. We start by understanding how frequently are computer jobs submitted to an industrial-scale data center and discover/explain two patterns with respect to the inter-arrival time (IAT) of job requests. Based on these, a novel generative process for modeling heterogeneous data is proposed for simulating job requests with the same statistical properties as the real data. On the computing platform side, we consider the problem of dynamic workload mapping in heterogeneous many-core systems via an efficient algorithm that maximizes performance under power constraints. While the generic mapping problem is NP-hard, we propose a close-to-optimal polynomial approach that can be used in an online scenario for heterogeneous workloads running on heterogeneous platforms.