paper-bot: paper-bot/2026-w28

[github-actions]

paper-bot batch

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

• 1803.09615 (522 cites, Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication 2018) — Homa: a receiver-driven low-latency transport protocol using network priorities
  • file: distributed-system/homa/1803.09615-homa-a-receiver-driven-low-latency-transport-protocol.md (be70091)
  • tl;dr: Homa is a receiver-driven transport protocol that uses in-network priority queues and controlled overcommitment to achieve exceptionally low latency for short messages while maintaining high throughput for large messages in datacenter networks.
• 1810.05934 (520 cites, Conference on Machine Learning and Systems 2018) — A System for Massively Parallel Hyperparameter Tuning
  • file: gpu-training/determined-asha/1810.05934-a-system-for-massively-parallel-hyperparameter-tuning.md (ab80e37)
  • tl;dr: Introduces ASHA, a simple and robust hyperparameter optimization algorithm that exploits parallelism and aggressive early-stopping, and describes its integration into Determined AI's production ML system.
• 93a06eb (508 cites, European Conference on Computer Systems 2018) — Optimus: an efficient dynamic resource scheduler for deep learning clusters
  • file: orchestration-scheduling/optimus/93a06eb-optimus-an-efficient-dynamic-resource-scheduler-for-deep.md (42d2ae7)
  • tl;dr: Optimus is a dynamic resource scheduler for deep learning clusters that uses online resource-performance models to allocate resources efficiently, reducing job completion times by up to 50% compared to static schedulers.

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.