Preliminary Performance Analysis of Multi-rail Fat-Tree Networks

摘要

Among the low-diameter, high-radix networks being deployed in next-generation HPC systems, dual-rail fat-tree networks are a promising approach. Adding additional injection connections (rails) to one or more network planes allows multi-rail fat-tree networks to alleviate communication bottlenecks. These multi-rail networks necessitate new design considerations, such as routing choices, job placements, and scalability of rails. We extend our fat-tree network model in the CODES parallel simulation framework to support multi-rail and multi-plane configurations in addition to different types of static routing, resulting in a powerful research vehicle for fat-tree network analysis. Our detailed packet-level simulations use communication traces from real applications to make performance predictions and to evaluate the impact of singleand multi-rail networks in conjunction with schemes for injection rail selection and intraplane routing.