Workshop: Workshop on Operational Data Analytics in HPC (HPC-ODA)
Authors: Rui Liu (Lawrence Berkeley National Laboratory), Paul Lin and Evan Weinberg (NVIDIA), Dhruva Kulkarni and Zhengji Zhao (Lawrence Berkeley National Laboratory), Ian Karlin (Queen's University), Garima Kochhar (NVIDIA), and Brian Austin (Lawrence Berkeley National Laboratory)
Abstract: The distribution of MPI message sizes issued by an application is a first-order determinant of its communication behavior, yet recovering this distribution typically requires application-level instrumentation and/or a priori knowledge of the communication pattern. We present an application-agnostic model of MPI communication that reconstructs the MPI message-size distribution from passive network telemetry: packet-level counters that modern NICs expose and that facility operators already collect continuously, system-wide, at negligible cost. The model requires no knowledge of the network topology, the communication peers, or the MPI operations in use, and can be applied retroactively to archived telemetry. Instantiated on the Cassini counters of NERSC's Perlmutter, the model estimates OSU Micro-Benchmark message counts within 0.5% and the dominant LAMMPS size classes within 10%, which is a first step toward instrumentation-free prediction of communication time.
Back to Workshop on Operational Data Analytics in HPC (HPC-ODA) Archive Listing
Back to Full Workshop Archive Listing