Blocking and Shared Memory Communication in OpenMPI and UCX

OpenMPI is a widespread implementation of the MPI Standard and uses UCX as its backend for communication. UCX supports both blocking and non-blocking communication modes, though OpenMPI only uses blocking by default to minimize communication latency.

My research explores the impact of changing communication modes on MPI performance by comparing the default blocking mode of communication with an alternative non-blocking approach. I used both micro-benchmarks and more holistic benchmarks that approximate real-life workloads to measure the effects on communication latency and energy consumption.

In this talk, I'll introduce this topic and then present my experimental setup, its results, and my conclusions.