Performance Evaluation of Data Transfer API for Rank Level Approximate Computing on HPC Systems
Yoshiyuki Morie, Yasutaka Wada, Ryohei Kobayashi, Ryuichi Sakamoto · 2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) · 2022
Approximate computing (AC) has attracted much attention to optimize tradeoffs among performance, power con-sumption, and computation results accuracy by adjusting data precision in applications. Even on HPC systems, AC is demanded to maximize performance under the limited power budget and hardware resources. To apply AC for HPC applications, we need to consider the character of each MPI rank in an application and optimize it with its appropriate data precision. However, we also need to perform data transfer while converting the precision of the target data. This paper proposes data pack/unpack APIs, which are applicable for standard MPI programs for HPC systems, for converting the data precision of the target data, and shows its performance evaluation. We can express data transfer among ranks with different data precision with the proposed APIs. The performance evaluation reveals the break-even point to apply AC for HPC applications from the perspective of data transfer volume.