Reproduce “Broadcast Rate Limits in Wi-Fi: A Forgotten Bottleneck for Collaborative Edge LLM Inference” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Liujianfu Wang, Yuyang Du, Shiqi Xu, Soung Chang Liew
The authors propose UDP broadcast for dispatching embeddings from a main node to expert workers in distributed mixture-of-experts inference, replacing serial sends with one transmission. Their method includes timeout-based TCP retransmission. In a wired test, Shadow-e4n16 UDP took 72% of NCCL’s expert-layer time. Wi-Fi broadcast was capped at 54 Mbps or lower; NS-3 simulation and emulation found an optimal 1,729 Mbps at 5 m, not demonstrated on commodity Wi-Fi.
Key insights
- UDP broadcast intrinsically matches the one-to-many communication pattern required by distributed MoE LLM inference, offering substantial speed advantages over sequential unicast methods.
- Current IEEE 802.11 Wi-Fi standards impose a 'forgotten bottleneck' of a legacy 54 Mbps cap on broadcast frame rates, regardless of the physical layer's actual capacity, severely impeding modern high-throughput data applications.
- NS-3 simulations reveal that optimal Wi-Fi broadcast rates for edge AI workloads are orders of magnitude higher (e.g., 3459 Mbps at 1m) than the current 54 Mbps ceiling, indicating significant untapped wireless potential.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.