Data transfer management policy optimization for unified memory

Abstract

OpenMP 4.5 refines the offload feature to better support heterogeneous computing. Explicit specification programming is currently required to optimize data transfer in OpenMP offload programs, but manual programming is not efficient or performant. Although DCU-supported unified memory provides a scheme for compilers to implicitly manage data transfers, target offload programs using unified memory perform poorly when program requests exceed the physical memory size. Therefore, this paper proposes a UMAT scheme for automatically optimizing unified memory-based data transfer management, which guides runtime management of data transfers by calculating the access frequency of data objects in the unified memory space. Test results show that the scheme has significant performance improvement for target offload programs using unified memory.