The vast supercomputing power at our disposal in the exascale era will go to waste unless we ensure that our applications can run at their peak performance. The amount of communication in an application will be the primary determinant of performance at those scales due to much faster cores but disproportionately slower networks. Fast, scalable and accurate modeling/simulation of the application’s communication is required to prepare parallel applications for exascale. Modeling and simulation have several use cases–performance prediction on a future architecture, network hardware co-design, and algorithmic research for future network topologies such as communication-avoiding algorithms, topology-aware job scheduling, task mapping and network routing protocols.
Nikhil Jain, Abhinav Bhatele, Michael P. Robson, Todd Gamblin, and Laxmikant V. Kale. Predicting application performance using supervised learning on communication features. In submission to Supercomputing 2013 (SC ‘13), 2013. LLNL-CONF-635857.