Study and evaluation of many-core CPU offloading for computing processes in environment adaptive software

Abstract

In recent years, the number of applications using diverse hardware such as FPGAs and GPUs has been increasing. However, utilizing these for high performance requires knowledge of hardware characteristics, which is quite difficult. Against this background, we propose environment-adaptive software that automatically converts and configures software code written by ordinary programmers for small-core CPUs according to the deployment environment, resulting in high-performance processing. This paper focuses on automatic offloading to many-core CPUs according to the computation type of processing, such as block matrix calculations and trivially parallel processing. We semantically analyze the existing application to be offloaded using an abstract syntax tree with pattern matching to determine whether the computation type has a replaceable OpenMP. If OpenMP is found, we replace it with that OpenMP and confirm the performance improvement. We confirm the effectiveness of the proposed method for automatic offloading by measuring processing time using an AMD Ryzen Threadripper 3995WX 64-core CPU.

This manuscript is a preprint and is currently under review at Journal of Network and Systems Management (Springer Nature).