Some New Approaches to MPI Implementations
Yuqing Xiong
TL;DR
The paper addresses MPI performance on exascale systems and proposes three complementary approaches: dynamically composable MPI libraries that tailor function sets to applications, frequency-based layering to reduce average stack depth, and a unified MPI-network/protocol design with per-function transport protocols and dedicated networks. Collectively, these strategies aim to boost performance by reducing unnecessary layers, enabling per-function optimizations, and integrating protocol and network design with MPI. While each approach offers potential gains, their practical realization will require substantial cross-layer engineering and careful integration. The proposed directions are designed to be combinable, offering a path toward significantly more efficient MPI implementations at scale.
Abstract
This paper provides some new approaches to MPI implementations to improve MPI performance. These approaches include dynamically composable libraries, reducing average layer numbers of MPI libraries, and a single entity of MPI-network, MPI-protocol, and MPI.