Reducing Cross-Pod Communication Overhead for MoE Model Training With Hybrid Parallelism in Multi-Tenant Clusters

Abstract

The massive parameter scale of sparsely-activated Mixture-of-Experts (MoE) models necessitates distributed training with hybrid parallelism. Placing such training tasks, i.e. mapping the logical partitions of an MoE model to available physical NPUs, is challenging. Due to the bandwidth and latency discrepancies between intra- and inter- Pods, the cross-Pod communication usually becomes a bottleneck. The high dispersion of NPUs in multi-tenant clusters exacerbates this issue further. However, a paucity of studies has considered the cross-Pod model placement problem. To address this challenge, we propose a novel model placement scheme tailored for MoE model training with hybrid parallelism in multi-tenant clusters. By quantifying the cross-Pod communication overhead incurred during MoE model training, the model placement is formulated as a 0-1 integer quadratic problem, which is NP hard. Motivated by the traffic difference between different parallelism, we decompose this problem into two subproblems. To solve the subproblems, we propose a lightweight two-stage algorithm based on Best-Fit strategy and neighborhood search. Experiments under different models and network topologies show that our model placement scheme can reduce cross-Pod traffic by 35.9% and cut communication time by 18.7% compared to state-of-the-art methods.