Research map / Architectures and efficiency / Mixture of experts
EasyBalance: Cross-Layer Load Balancing in Distributed MoE Inference
Yize Wu, KE GAO, Ling Li, Yanjun Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add EasyBalance cross-layer scheduling | Qwen3-30B · 2wikimqa EP=8 | GPU under-utilization: 0.37 -> 0.21 | Improved |
| add EasyBalance cross-layer scheduling | Moonlight-16B · repobench-p EP=8 | GPU under-utilization: 0.46 -> 0.29 | Improved |
| add EasyBalance cross-layer scheduling | Qwen3-30B · trec EP=2 | GPU under-utilization: 0.08 -> 0.04 | Improved |
| add micro-batching (4 splits) (ablation) | Qwen3-30B · 2wikimqa EP=8 | end-to-end latency (s): 10.15 -> 8.60 | Improved |
| add micro-batching (4 splits) (ablation) | Moonlight-16B · 2wikimqa EP=8 | end-to-end latency (s): 5.06 -> 4.40 | Improved |
| use MaxUtil instead of linear heuristics CumUtil/DiffPeak (ablation) | EasyBalance · 2wikimqa Qwen3-30B | end-to-end latency (s): 8.27 -> 8.21 | Improved |
| use cheaper DiffPeak scheduler instead of MaxUtil (ablation) | EasyBalance · 2wikimqa Qwen3-30B | scheduling latency per step (ms): 0.277 -> 0.130 | Improved |
| add EasyBalance on top of EPLB(128) (ablation) | EPLB · 2wikimqa Qwen3-30B | GPU under-utilization: 0.23 -> 0.14 | Improved |
| add EasyBalance on top of EPLB(136) (ablation) | EPLB with replicas · trec Qwen3-30B | GPU under-utilization: 0.28 -> 0.17 | Improved |
| add 8 redundant experts to EPLB (128 to 136 physical) (ablation) | EPLB · 2wikimqa Qwen3-30B | GPU under-utilization: 0.23 -> 0.22 | No clear effect |
| increase expert-parallel size from 2 to 8 without EasyBalance (ablation) | Qwen3-30B · 2wikimqa | GPU under-utilization: 0.10 -> 0.37 | Hurt |
| use very small micro-batch count (2) (ablation) | EasyBalance · LongBench | imbalance mitigation | Hurt |
Closest papers by meaning
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.91)
- Mixture of Lookup Experts (ICML 2025 · similarity 0.90)
- Shortcut-connected Expert Parallelism for Accelerating Mixture of Experts (ICML 2025 · similarity 0.89)
- Occult: Optimizing Collaborative Communications across Experts for Accelerated Parallel MoE Training and Inference (ICML 2025 · similarity 0.89)
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · similarity 0.89)
- Toward Efficient Inference for Mixture of Experts (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
