High Throughput Large Scale Sorting on a CPU-FPGA Heterogeneous Platform

Recently accelerating sorting using FPGA has been of growing interest in both industry and academia. However, the supported size of data set is usually small for FPGA-only sorting designs due to limited on-chip memory. In this paper, we propose a design to speed-up large scale sorting using a CPU-FPGA heterogeneous platform. We first optimize a fully-pipelined merge sort based accelerator and employ several such designs working in parallel on FPGA. The partial results from the FPGA are then merged on the CPU. On the Intel QuickAssist QPI FPGA Platform, for a range of data set size, we improve the throughput by 2.9× and 1.9× compared with CPU-only and FPGA-only baselines, respectively. Compared with the state-of-the-art FPGA implementation for sorting, our design achieves 2.3× throughput improvement.