Scaling Multi-Node Mixture-of-Experts Inference Using Expert Activation Patterns

TL;DR

This paper analyzes challenges in multi-node MoE inference for large language models, profiling expert activation patterns and proposing workload-aware strategies to reduce communication overhead and improve efficiency.

Contribution

It systematically characterizes expert activation properties and introduces workload-aware micro-batch grouping and expert placement strategies to optimize multi-node MoE inference.

Findings

Profiling reveals persistent expert load imbalance and domain-specific activation patterns.

Proposed strategies reduce inter-node communication by up to 20x.

Optimizations lead to lower latency and better accelerator utilization.

Abstract

Most recent state-of-the-art (SOTA) large language models (LLMs) use Mixture-of-Experts (MoE) architectures to scale model capacity without proportional per-token compute, enabling higher-quality outputs at manageable serving costs. However, MoE inference at scale is fundamentally bottlenecked by expert load imbalance and inefficient token routing, especially in multi-node deployments where tokens are not guaranteed to be routed to local experts, resulting in significant inter-node all-to-all communication overhead. To systematically characterize these challenges, we profile SOTA open-source MoE models, including Llama 4 Maverick, DeepSeek V3-671B, and Qwen3-230B-A22B, on various datasets and collected over 100k real expert activation traces. Upon studying the expert activation patterns, we uncover various persistent properties across all the frontier MoE models: variable expert load imbalance, domain-specific expert activation where expert popularity shifts across task families (code, math, chat, general), and a strong correlation between prefill and decode expert activations. Motivated by these findings, we propose workload-aware micro-batch grouping and an expert placement strategy to maximize token locality to the destination expert, thereby reducing inter-node communication. Across models and datasets, these optimizations help reduce all2all communication data up to 20, resulting in lower MoE decode latency and better accelerator utilization.