Efficient Data Processing using On-the-Fly Host-PIM Interactions in a Commodity PIM System

초록

Processing-in-Memory (PIM) systems provide a highly parallel data processing platform for data-intensive applications by integrating a processing unit (PU) with each DRAM bank. To exploit the system's parallel computing capacity, applications need to divide the input data into partitions so that PUs can process them in parallel. However, such applications often impose different loads across PUs. For example, in a Sparse Matrix-Vector Multiplication (SpMV) application, the computing time per PU varies widely with the number of nonzero elements in the assigned data partition. Unbalanced loads incur the notorious straggler problem in parallel systems, including the PIM system. Hence, a precise data partitioning scheme is desired in the applications, which is a daunting task due to the varying workload characteristics depending on the input data. In this paper, we propose an On-the-Fly host-PIM Interaction Scheme (OFIS) that solves the problem by enabling on-demand data assignment to each (or a group of) PU that has completed processing assigned data earlier than others, which is not possible in the current commercial PIM system due to its collective management of PUs. We implement OFIS in the commodity PIM system and evaluate its performance using the SpMV, PageRank, and regular expression matching applications. Experimental results show that the three applications, without employing any precise data partitioning scheme for load balancing, achieve speedups of up to 9.7, 1.4, and 5.4, respectively, using the proposed scheme.

키워드