Exploring GPU-to-GPU Communication: Insights into Supercomputer Interconnects

TL;DR

This paper characterizes GPU-to-GPU communication in supercomputers, revealing untapped bandwidth and optimization opportunities across different architectures to improve performance and efficiency.

Contribution

It provides a comprehensive performance analysis of intra-node and inter-node GPU interconnects on three supercomputers, offering practical insights for optimization.

Findings

Untapped bandwidth exists in current GPU interconnects.

Performance bottlenecks vary across architectures.

Opportunities for software and network optimization are significant.

Abstract

Multi-GPU nodes are increasingly common in the rapidly evolving landscape of exascale supercomputers. On these systems, GPUs on the same node are connected through dedicated networks, with bandwidths up to a few terabits per second. However, gauging performance expectations and maximizing system efficiency is challenging due to different technologies, design options, and software layers. This paper comprehensively characterizes three supercomputers - Alps, Leonardo, and LUMI - each with a unique architecture and design. We focus on performance evaluation of intra-node and inter-node interconnects on up to 4096 GPUs, using a mix of intra-node and inter-node benchmarks. By analyzing its limitations and opportunities, we aim to offer practical guidance to researchers, system architects, and software developers dealing with multi-GPU supercomputing. Our results show that there is untapped bandwidth, and there are still many opportunities for optimization, ranging from network to software optimization.