Performance and Energy Aware Workload Partitioning on Heterogeneous Platforms

Heterogeneous platforms which employ a mix of CPUs and accelerators such as GPUs have been widely used in the high-performance computing area [1]. 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 heterogeneous platforms. The main difficulty is that the processors in heterogeneous systems usually feature distinct characteristics, which is not presented in homogeneous platforms. This difficulty brings two main challenges that prevent achieving the promised performance and energy efficiency. One main challenge is the efficient utilization of the different types of processors in heterogeneous platforms. Many studies [2], [3], [4] have been done to increase the processor utilization of heterogeneous platforms. However, this type of work assumes that workload has already been implemented or requires online profiling information. The second main challenge is the large cost in design time to partition workload for heterogeneous platform. Numerous efforts [5], [6], [7] have been made in automatic workload partitioning on heterogeneous platforms. However, these efforts only consider workload partitioning via data partitioning (DP). Workload here refers to the amount of computation (measured by flops) and memory traffic (measured by bytes) to be executed at the algorithmic level. Judicious code partitioning (CP) between distinct processors has been shown to be able to achieve better performance/energy than DP for running workloads on heterogenous platforms [8]. An example is shown in Figures 1 and 2. The CP-based implementation maps the vector addition and power operations in all iterations onto the CPU and GPU, respectively. The DP-based implementation maps the iterations (i.e., dataset size) evenly onto the CPU and GPU based on the CPU and GPU performance. The results of running these two implementations on two heterogeneous platforms indicate that CP can have better performance than DP but using different combinations of CPU and GPU also impacts this observation. To help developers consider both DP and CP for workload partitioning in design time with affordable cost, my thesis research aims to develop a lightweight tool to help developers partition workload and select appropriate workload partition (WP).