Eliminating Hidden Serialization in Multi-Node Megakernel Communication
Publish this paper in The AIPR Journal
Are you an author? Turn this AI review into a permanent, citable journal entry with a cover, open comments, and Scholar metadata.
AIPR assessment
This is a hard, crowded systems problem, with many groups optimizing MoE communication, megakernels, and GPU-initiated RDMA at the same time. The strengths reinforce each other well: a sharp bottleneck diagnosis, a small mechanism with a clear causal story, and strong hardware-backed evaluation. The main weaknesses interact less severely than in many systems papers, but they still matter: the artifact is not openly packaged, the transport coverage is broad but not exhaustive, and the generality
Abstract
Recent megakernel designs for Mixture-of-Experts (MoE) inference fuse expert computation with fine-grained, GPU-initiated communication into a single persistent GPU kernel, and outperform collective-based MoE on a single node by overlapping data transfer with compute at tile granularity. This benefit does not carry over cleanly to multi-node inference, where experts span many nodes connected by an RDMA fabric. Communication-bound MoE models regress by up to $10\times$ on 8 nodes, and the regression worsens with node count. We trace this regression to hidden serialization in proxy-based RDMA transports. The ordering requirement between each tile transfer and its completion signal forces a fence that drains the NIC pipeline, and its cost grows with the number of concurrent transfers. As a result, models whose per-expert compute is too small to absorb this inflated network latency expose communication on the critical path. We present \emph{Perseus}, which eliminates this serialization through two techniques. \emph{Decoupled signaling} batches fences at per-destination granularity, reducing fence count by $8\times$. \emph{NIC-side ordering} replaces proxy stalls with hardware fence flags, so the proxy never blocks. On proxy-based transports, Perseus achieves up to 10.3$\times$ end-to-end speedup. Perseus on IBRC matches or exceeds IBGDA GPU-direct by up to 1.2$\times$, which shows that serialization, rather than the choice between proxy-based and GPU-direct transport, is what bounds multi-node megakernel performance.
Score Breakdown
More from this week
- Beyond Visual Fidelity: Benchmarking Super-Resolution Models for Large-Scale Remote Sensing Imagery via Downstream Task Integration
- Succinct Graph Representations and Algorithmic Applications
- A Faster Deterministic Algorithm for Fully Dynamic Maximal Matching
- Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs
- Claw-Eval-Live: A Live Agent Benchmark for Evolving Real-World Workflows