Reducing the overhead of an MPI application-level migration approach

• Solution to reduce memory and I/O overhead in a checkpoint-based migration tool.
• The proposal splits the checkpoint files in small chunks and overlaps I/O phases.
• Results prove its efficiency, both in terms of memory consumption and I/O times.
• Reduction in memory usage enables migration in applications with large state files.

Process migration provides many benefits for parallel environments including dynamic load balance, data access locality, or fault tolerance. This work proposes a solution that reduces the memory and I/O overhead in an application-level checkpoint-based migration approach. The proposal splits the checkpoint files in order to overlap the writing of the state in the terminating processes with the read and restarting operation in the newly spawned processes. It has been tested using the MPI NAS Parallel Benchmarks, showing encouraging results, both in terms of memory consumption and I/O migration times.