PRIME: Pseudo-Random Integrated Multi-Part Entropy for Adaptive Packet Spraying in AI/ML Data centers

Abstract

Large-scale distributed training in production data centers place significant demands on network infrastructure. In particular, significant load balancing challenges arise when processing AI/ML workloads, consisting of low-entropy, bursty and long-lived flows. Existing solutions designed for Ethernet, such as Equal-Cost Multi-Path (ECMP) struggle to maintain high network utilization. While major industry players (e.g., Ultra Ethernet Consortium) and parts of academia have proposed packet spraying to enhance AI/ML workload performance, we argue that existing packet spraying solutions lead to buffer inflation over time, negatively affecting network performance. Specifically, when ACK coalescing is used, these solutions lead to stale information, degrading network performance. Additionally, in asymmetric network conditions- such as mix of ordered an unordered traffic, or link degradation and failures- existing packet spraying solutions often lead to increased tail latency. In this paper, we present the design and evaluation of PRIME, a pseudo-randomized round-robin approach to packet spraying that considers the network topology to optimize load distribution and performance. PRIME uses congestion as an indicator to re-balance the load. To this extent, PRIME takes into account various congestion signals, accounting for congestion severity, and their decay times to avoid network hotspots. We extensively evaluated PRIME using large-scale production-level simulator. Our results indicate that, compared to existing solutions, PRIME leads to up to 15% improvement for permutation traffic and up to 27% improvement in network degradation scenarios