The Impact of Adopting Computational Storage in Heterogeneous Computing Systems

Abstract

Heterogeneous computing systems have been the focus of major efforts in pursuit of Exascale computing over the past few years. These heterogeneous computing systems, often contain various accelerators, can significantly improve the computation performance thanks to their parallel compute cores. Much research has exploited accelerators to reduce the computation time; however, I/O access time or end-to-end total execution time is often overlooked. Typically, when I/O access time with the storage is considered, the performance gain from accelerators can be substantially degraded. In this work, we designed a computational storage in a heterogeneous computing system. Our evaluation employs a real-world compute-bound machine learning application on three computing systems: A conventional CPU system, a modern FPGA system, and the computational storage system. Compared to CPU system, we found that the modern FPGA system can achieve a 100x speedup for computation and 62x better energy efficiency, but only 26 × speedup and 16 × energy efficiency improvement when data movement time is taken into account; whereas the computational storage system can achieve 63 × computation speedup and 51 × better energy efficiency including data movement time. The experiment demonstrates that a compute-bound problem on conventional CPU system may become an I/O-bound problem on heterogeneous system. By adopting computational storage in heterogeneous system, we find that both the computation performance and the I/O performance is improved.