Dice Question Streamline Icon: https://streamlinehq.com

Runtime cost of MPI gather/scatter for tailored datatypes

Evaluate the runtime overhead of MPI gather and scatter operations when exchanging subsets of struct attributes via user-defined derived MPI datatypes.

Information Square Streamline Icon: https://streamlinehq.com

Background

The paper introduces annotations to generate MPI datatypes that exchange only selected struct attributes, improving communication efficiency by avoiding unnecessary bytes.

Selecting subsets of attributes can cause MPI to gather/scatter noncontiguous data from memory, and the authors state that the cost of these operations is presently unknown.

References

Picking individual attributes that are to be exchanged from a struct or an array of structs means that MPI has to gather and scatter data from the memory. It is not clear how expensive these steps are.

— An extension of C++ with memory-centric specifications for HPC to reduce memory footprints and streamline MPI development (2406.06095 - Radtke et al., 10 Jun 2024) in Section 3.3 (MPI datatypes over structs), Impact paragraph