A Lightweight Communication Interface for Asynchronous Many-Task Systems

MPI underpins communications in a wide swath of HPC applications. These have functioned very well for the programming paradigms that they were primarily designed for, namely bulk-synchronous applications using explicit message-passing for communication. However, asynchronous many-task runtimes present different communication characteristics than most traditional HPC applications; these communication interfaces are not well-optimized and do not serve the needs and requirements of AMT runtimes. We present an overview of the design of the Lightweight Communication Interface (LCI), a communication library and research tool developed at the University of Illinois to study communication interface and design decisions in the context of multithreaded communication and dynamic compute frameworks, and demonstrate how it can be used to improve the performance of AMT runtimes. We showcase current work with the HPX and PaRSEC frameworks and compare results with a set of task-based latency and bandwidth benchmarks.