MPICkpt: A Transparent Checkpointing Tool for MPI

Abstract

This work designs and implements MPICkpt, a transparent and multifunctional checkpointing system for MPI. In addition to employing two kinds of checkpointing techniques, i.e. coordinated and independent. MPICkpt includes various optimized checkpointing functions, e.g. forded, incremental, and compressed functions, to reduce the checkpointing overhead. The user can either take checkpoints in some specified locations of the source program or let the system do so automatically and in a fully transparent manner. In addition, a checkpointing visualization facility is developed and implemented to evaluate the effectiveness of various checkpointing techniques. Experimental results indicate that the performance of checkpointing depends on how the applications function.