Efficient checkpoint/Restart of CUDA applications

We present NVCR which enables transparent checkpoint and restart of CUDA applications. NVCR, works as an extension of major system-level checkpoint software such as BLCR and DMTCP, employs proxy-process and application accesses GPU devices via the proxy-process to improve the compatibility with latest CUDA runtime software. To reduce the overhead of inter-process communications, NVCR efficiently uses SYSV IPC shared memory as CUDA pinned memory. Performance evaluations using micro benchmarks and Amber as a real application show that NVCR' overhead is acceptably low.