Tackling the Data-Parallel Load Balancing Bottleneck in LLM Serving: Practical Online Routing at Scale

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Bu, Tianci, Lyu, Yuan, Chen, Zixi, Song, Chendong, Liang, Hong, Gurung, Tsepten, Fan, Yuwei, Ye, Yinyu, Zhou, Zijie
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866918488945197056
author Bu, Tianci
Lyu, Yuan
Chen, Zixi
Song, Chendong
Liang, Hong
Gurung, Tsepten
Fan, Yuwei
Ye, Yinyu
Zhou, Zijie
author_facet Bu, Tianci
Lyu, Yuan
Chen, Zixi
Song, Chendong
Liang, Hong
Gurung, Tsepten
Fan, Yuwei
Ye, Yinyu
Zhou, Zijie
contents Data-parallel (DP) load balancing has emerged as a first-order bottleneck in large-scale LLM serving. When a model is sharded across devices via tensor parallelism (TP) or expert parallelism (EP) and replicated across many DP workers, every decode step ends in a synchronization barrier whose latency is set by the most heavily loaded worker; even modest persistent imbalance across DP workers compounds, step after step, into a substantial fraction of wasted compute. The problem is hard for reasons specific to LLM decoding: assignments are sticky (migrating KV caches has a high cost), per-request loads grow over time, arrivals are non-stationary, and the router must decide within a sub-100\,ms decode budget over hundreds of waiting requests and tens of workers. We present \textbf{BalanceRoute}, a family of practical online routing algorithms that target this bottleneck. The first, \textbf{BR-0}, requires no prediction infrastructure and uses a piecewise-linear F-score that captures the sharp asymmetry between admissions that fill safe margin and those that overflow into the envelope; a two-stage decomposition keeps per-step cost compatible with millisecond-scale scheduling. The second, \textbf{BR-H}, generalizes BR-0 with a short, constant lookahead $H$ and a lightweight termination-classifier interface, extending the F-score to a horizon-discounted form. We deploy BalanceRoute on a 144-NPU cluster and evaluate against vLLM baselines on both a proprietary production trace and the public Azure-2024 trace. Across both workloads, BalanceRoute substantially reduces average DP imbalance and improves end-to-end serving throughput.
format Preprint
id arxiv_https___arxiv_org_abs_2605_06113
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Tackling the Data-Parallel Load Balancing Bottleneck in LLM Serving: Practical Online Routing at Scale
Bu, Tianci
Lyu, Yuan
Chen, Zixi
Song, Chendong
Liang, Hong
Gurung, Tsepten
Fan, Yuwei
Ye, Yinyu
Zhou, Zijie
Distributed, Parallel, and Cluster Computing
Data-parallel (DP) load balancing has emerged as a first-order bottleneck in large-scale LLM serving. When a model is sharded across devices via tensor parallelism (TP) or expert parallelism (EP) and replicated across many DP workers, every decode step ends in a synchronization barrier whose latency is set by the most heavily loaded worker; even modest persistent imbalance across DP workers compounds, step after step, into a substantial fraction of wasted compute. The problem is hard for reasons specific to LLM decoding: assignments are sticky (migrating KV caches has a high cost), per-request loads grow over time, arrivals are non-stationary, and the router must decide within a sub-100\,ms decode budget over hundreds of waiting requests and tens of workers. We present \textbf{BalanceRoute}, a family of practical online routing algorithms that target this bottleneck. The first, \textbf{BR-0}, requires no prediction infrastructure and uses a piecewise-linear F-score that captures the sharp asymmetry between admissions that fill safe margin and those that overflow into the envelope; a two-stage decomposition keeps per-step cost compatible with millisecond-scale scheduling. The second, \textbf{BR-H}, generalizes BR-0 with a short, constant lookahead $H$ and a lightweight termination-classifier interface, extending the F-score to a horizon-discounted form. We deploy BalanceRoute on a 144-NPU cluster and evaluate against vLLM baselines on both a proprietary production trace and the public Azure-2024 trace. Across both workloads, BalanceRoute substantially reduces average DP imbalance and improves end-to-end serving throughput.
title Tackling the Data-Parallel Load Balancing Bottleneck in LLM Serving: Practical Online Routing at Scale
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2605.06113