AutoMatch: An automated framework for relative performance estimation and workload distribution on heterogeneous HPC systems
Ahmed E. Helal, Wu-chun Feng, Changhee Jung, Yasser Y. Hanafy · 2017
Porting sequential applications to heterogeneous HPC systems requires extensive software and hardware expertise to estimate the potential speedup and to efficiently use the available compute resources in such systems. To streamline this daunting process, researchers have proposed several “black-box” performance prediction approaches that rely on the performance of a training set of parallel applications. However, due to the lack of a diverse set of applications along with their optimized parallel implementations for each architecture type, the predicted speedup by these approaches is not the speedup upper-bound, and even worse it can be misleading, if the reference parallel implementations are not equally-optimized for every target architecture. This paper presents AutoMatch, an automated framework for matching of compute kernels to heterogeneous HPC architectures. AutoMatch uses hybrid (static and dynamic) analysis to find the best dependency-preserving parallel schedule of a given sequential code. The resulting operations schedule serves as a basis to construct a cost function of the optimized parallel execution of the sequential code on heterogeneous HPC nodes. Since such a cost function informs the user and runtime system about the relative execution cost across the different hardware devices within HPC nodes, AutoMatch enables efficient runtime workload distribution that simultaneously utilizes all the available devices in performance-proportional way. For a set of open-source HPC applications with different characteristics, AutoMatch turns out to be very effective, identifying the speedup upper-bound of sequential applications and how close the parallel implementation is to the best parallel performance across five different HPC architectures. Furthermore, AutoMatch's workload distribution scheme achieves approximately 90% of the performance of a profiling-driven oracle.