Deterministic adaptive routing through hashing in CXL-based memory fabrics

Abstract

however, the routing algorithm must be carefully designed to fully exploit the path diversity to provide high throughput. Different routing algorithms have been proposed, with many based on randomization when selecting an output port, including many adaptive routing schemes. While they can successfully balance the traffic, they suffer from instantaneous contention and their performances are limited. On the other hand, Deterministic routing has performance trade-offs for different traffics, while it works best in balanced traffic, it performs poorly for permutation traffic. In this work, we propose a novel routing algorithm that leverages the use of hashing and adaptive routing to deterministically and adaptively distribute packets across the network with minimal collision. Our evaluation shows that our proposal outperforms other routing algorithms in most configurations and evaluation metrics. Specifically, it can provide up to 30\% runtime improvement compared to Random and Adaptive Routing, and even matches the optimal performance of Deterministic in Uniform Random traffic. In addition, we also evaluated the impacts of communicating with large message sizes and the impacts of different packet sizes on performance.

Recently, the development of Compute Express Link (CXL) has enabled heterogeneous computing with a large number of components, and it is important to make sure that communication will not be the system's bottleneck. Fat-tree is a popular option thanks to its high path diversity