Checkpointing Tools in a Supercomputer Center
G. I. Savin, Борис Михайлович Шабанов, R. S. Fedorov, Anton V. Baranov, Pavel Telegin · Lobachevskii Journal of Mathematics · 2020
Abstract The article describes the problem of automatic checkpoints creation/data restoration for the jobs that run on a single supercomputer node. The paper formulates the requirements for the checkpointing and restore tools in the supercomputer job management system. Berkeley Lab Checkpoint/Restart (BLCR), Checkpoint Restore In Userspace (CRIU), and Distributed MultiThreaded Checkpointing (DMTCP) tools are examined. It is shown that the DMTCP tool better meets the stated requirements. Experimental estimates of computational performance and impact on efficiency for DMTCP are presented. The problems of checkpointing tools’ integration into the SUPPZ job management system used at the JSCC RAS are considered. Recommendations on the practical use of automatic checkpoint/restore tools are given.