Abstract
Collective communications occupy 20-90% of total execution times in many MPI applications. In this paper, we propose strategies for automatically identifying the most time-consuming collective operations that also act as scalability bottlenecks. We then explore the use of MPI-3’s non-blocking collectives for these communications. We also rearrange the codes to adequately overlap the independent computations with the non-blocking collective communications. Applying these strategies for different graph and machine learning applications, we obtained up to 33% performance improvements for large-scale runs on a Cray supercomputer.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
