Checkpointing improvement in ParJava environment

摘要

The prototype of checkpoint recovery protocol is implemented in current version of ParJava environment. Coordinated use of checkpointing tools can decrease development and debugging efforts of application programmer. Proposed approach to arrangement of checkpoints allows to keep checkpointing overheads small enough. ParJava is installed on several clusters: ISP RAS, Joint Supercomputer Center (Moscow), and Moscow University. Presently some applications (neural networks, flood modeling) are being developed using ParJava.