Performance and Energy Aware Workload Partitioning on Heterogeneous Platforms

Heterogeneous platforms which employ a mix of CPUs and GPUs have been widely used in many different areas such as embedded computing and high performance computing. Such heterogeneous platforms have the potential to offer higher performance at lower energy cost than homogeneous platforms. However, it is rather challenging to actually achieve the high performance and energy efficiency promised by many heterogeneous platforms. One reason is that a straightforward division of the workload of an application among the processors usually cannot efficiently utilize the distinct processors of heterogeneous platforms. Another reason is that a heterogeneous platform presents a large design space for workload partitioning among different processors.

To help application developers explore workload partitions (WPs) on heterogeneous platforms for exploiting performance and energy potential, we make four contributions in this dissertation. First, as a case study, we conduct an in-depth investigation of different strategies of partitioning the workload for the data assembly (DA) stage in finite element method applications. By running different WPs of DA on a broad range of heterogeneous hardware platforms, we examine the performance and energy impacts of workload partitioning on heterogeneous platforms. Second, we develop a performance model, PerDome, to estimate the performance potential of different WPs on heterogeneous platforms. Third, we build a framework, PeaPaw, to assist application developers to find a WP that has high potential leading to high performance or energy efficiency before actual implementation. The PeaPaw framework includes both analytical performance/energy models and two sets of workload partitioning guidelines. Based on the design goal (i.e., performance or energy), application developers can obtain a workload partitioning guideline and use PeaPaw to estimate the performance or energy of designed WPs on a given heterogeneous platform. Last, we enhance the capability of PeaPaw in handling more complicated application memory behaviors and integrated heterogeneous platforms. Our results show that PaPaw+ can provide higher accuracy of WP performance estimation and more effective workload partitioning guidelines on both discrete and integrated heterogeneous platforms than PeaPaw. With these contributions, this dissertation not only advances the understanding of interplay between workload partition and performance/energy of heterogeneous platforms but also provides practical techniques to guide efficient workload partitioning on heterogeneous platforms.