A protocol reconfiguration and optimization system for MPI

Modern high performance computing (HPC) applications, for example adaptive mesh refinement and multi-physics codes, have dynamic communication characteristics which result in poor performance on current Message Passing Interface (MPI) implementations. The degraded application performance can be attributed to a mismatch between changing application requirements and static communication library functionality. To improve the performance of these applications, MPI libraries should change their protocol functionality in response to changing application requirements, and tailor their functionality to take advantage of hardware capabilities. This dissertation describes Protocol Reconfiguration and Optimization system for MPI (PRO-MPI), a framework for constructing profile-driven reconfigurable MPI libraries; these libraries use past application characteristics (profiles) to dynamically change their functionality to match the changing application requirements. The framework addresses the challenges of designing and implementing the reconfigurable MPI libraries, which include collecting and reasoning about application characteristics to drive the protocol reconfiguration and defining abstractions required for implementing these reconfigurations. Two prototype reconfigurable MPI implementations based on the framework - Open PRO-MPI and Cactus PRO-MPI - are also presented to demonstrate the utility of the framework. To demonstrate the effectiveness of reconfigurable MPI libraries, this dissertation presents experimental results to show the impact of using these libraries on the application performance. The results show that PRO-MPI improves the performance of important HPC applications and benchmarks. They also show that HyperCLaw performance improves by approximately 22% when exact profiles are available, and HyperCLaw performance improves by approximately 18% when only approximate profiles are available.Sandia National Laboratories; Department of EnergyComputer ScienceDoctoralUniversity of New Mexico. Dept. of Computer ScienceBridges, PatrickGhani, NasirArnold, DorianRiesen, Rol