paper-bot: paper-bot/2026-w26

[github-actions]

paper-bot batch

4 papers from top systems venues, ranked by citation count. Each paper is its own commit so reviews are surgical.

• 2403.02310 (580 cites, USENIX Symposium on Operating Systems Design and Implementation 2024) — Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve
  • file: gpu-training/microsoft-sarathi-serve/2403.02310-taming-throughput-latency-tradeoff-in-llm-inference-with.md (edf50cf)
  • tl;dr: Sarathi-Serve is an LLM inference scheduler that uses chunked-prefills and stall-free scheduling to simultaneously achieve high throughput and low latency, significantly outperforming vLLM across various models and hardware configurations.
• 2006.06762 (565 cites, USENIX Symposium on Operating Systems Design and Implementation 2020) — Ansor : Generating High-Performance Tensor Programs for Deep Learning
  • file: gpu-training/ansor/2006.06762-ansor-generating-high-performance-tensor-programs-for-deep.md (ee71576)
  • tl;dr: Ansor is a tensor program generation framework that uses hierarchical search space sampling, evolutionary search, and a learned cost model to automatically generate high-performance tensor programs for deep learning on various hardware platforms.
• 1903.06701 (532 cites, Symposium on Networked Systems Design and Implementation 2019) — Scaling Distributed Machine Learning with In-Network Aggregation
  • file: gpu-training/switchml/1903.06701-scaling-distributed-machine-learning-with-in-network-aggregation.md (3ee01a1)
  • tl;dr: SwitchML accelerates distributed ML training by performing in-network aggregation of model updates using programmable switch dataplanes, reducing communication volume and speeding up training by up to 300%.
• ea639e0 (519 cites, Symposium on Networked Systems Design and Implementation 2019) — Monoxide: Scale out Blockchains with Asynchronous Consensus Zones
  • file: distributed-system/monoxide/ea639e0-monoxide-scale-out-blockchains-with-asynchronous-consensus-zones.md (a4b262e)
  • tl;dr: Monoxide scales blockchain throughput linearly with the number of mining nodes by partitioning the network into independent consensus zones that process transactions asynchronously, while ensuring cross-zone atomicity via a novel relay transaction mechanism.

How to review

Comment on this PR with one of:

• /refine <id> <free-form instructions> — re-write the analysis with your feedback. Folds into the original commit.
• /regenerate <id> — re-fetch the paper and run the full analyzer again.
• /reject <id> — revert the paper's commit.

<id> can be an arXiv id (when shown), the 7-char paper-id, or the file path. Examples:

/refine 2401.09670 把 evaluation 段补充和 vLLM 的对比数字
/regenerate 72f77a3
/reject gpu-training/misc/2024-some-paper.md

When done, merge with Squash & merge to keep main history clean.