Assessing the Impact of Network Compression on Molecular Dynamics and Finite Element Methods

摘要

Efficient communication in parallel applications is one of the main challenges for the scalability of supercomputers, both in weak and strong scaling environments. In the past, several compression techniques have been proposed as a way to improve the performance and scalability of parallel applications. Those works have shown significant speed-ups when applying compressors to the MPI transfers of certain algorithmic kernels. However, these techniques have not seen widespread adoption in current supercomputers. This paper evaluates the bottlenecks of network compression that have precluded their generalized adoption in HPC environments. In order to evaluate their impact on real applications we integrated multiple MPI compression schemes into two production applications: a computational mechanics code dominated by point-to-point communication, and a molecular dynamics code dominated by collective communications. While the applications observe some improvements when applying aggressive lossy compression schemes on systems ranging from 4 to 256 processors, the overall results seem to contradict earlier research. We conclude that HPC data traffic tends to be too statistically random to be captured by general lossless compressors, and that the size of MPI message is most often not the limiting component of these communication-bound applications. We also observed that aggressive lossy compression worked well and did not distort the results of the evaluated applications. This suggests that reducing network bandwidth in conjunction with message compression may be an interesting technique to increase energy efficiency in HPC systems.