DCU-CHK: checkpointing for large-scale CPU-DCU heterogeneous computing systems

By utilizing the superior computing power of accelerators, heterogeneous architectures have become increasingly popular in high-performance computing (HPC) systems. Meanwhile, the scale of HPC systems also continuously increases, which poses challenges to resilience. The Hygon DCU, a domestic-developed accelerator, has been used in a growing number of Chinese-made supercomputers. Therefore, it is crucial to provide checkpointing support for the CPU-DCU platform. This paper proposes DCU-CHK, a novel checkpointing scheme for large-scale CPU-DCU heterogeneous computing systems. The scheme provides transparent checkpointing for HIP applications and employs an address translation mechanism to ensure the robustness of restarting. The scheme is implemented based on DMTCP. Experimental results demonstrate the effectiveness and scalability of the scheme.