ENERGY-TIME PERFORMANCE OF HETEROGENEOUS COMPUTING SYSTEMS: MODELS AND ANALYSIS

Abstract

With heterogeneity becoming the norm in most computing platforms today, one of the key challenges is to determine the set of energy-time efficient system configurations among the large system configuration space to execute a parallel application. This thesis addresses this challenge using a measurement-driven analytical model that determines both time and energy efficient system configurations. Based on our taxonomy of heterogeneity, we first propose a core analytical model for a baseline inter-node heterogeneous system consisting of brawny and wimpy nodes. The key novelties include modeling both inter and intra-node resource overlaps and resource contention. The scalability of our core model is shown by extending it with inter- and intra-core contention for a Many Integrated Core architecture system, and inter- and intra-node communication for hybrid programs. The core model and its extensions are applied to determine energy-time efficient system configurations and expose a number of insights. Firstly, there are multiple Pareto-optimal configurations that can be approximated using a distinct energy-deadline Pareto-frontier. Secondly, our energy proportionality analysis reveals that inter-node heterogeneous clusters enable the scaling of the energy-proportionality wall by exposing sub-linear energy-proportional configurations. Lastly, we introduce a new metric called useful-computation-ratio which can be used to further optimize the Pareto-frontier.