Reproduce “Load Balancing with Evolving Workloads in Large-Scale Language Model Serving” 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

Zixi Chen, Tianci Bu, Chendong Song, Xin Lu, Yinyu Ye, Zijie Zhou

The authors propose BalanceRoute, which jointly admits waiting requests and places them using workers’ cached-token workloads rather than request counts. It minimizes aggregate imbalance, with optional short-lookahead predictions of completions. On proprietary trace 30002, BR-0 achieved lower mean imbalance, lower decode-stage TPOT P95, and higher throughput than FCFS+Count-JSQ. The theorem concerns exact optimization under stated assumptions, not practical-router guarantees.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.