---
格式版本: 2
标题: "Benchmarking LLM-Guided Control-Plane Policies for Backend Fault Isolation in HAProxy"
原文链接: "https://arxiv.org/abs/2608.10532"
发布日期: "2026-08-11"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Tue, 11 Aug 2026 06:15:20 UTC (813 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-12T11:23:31+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-12T11:20:34+08:00"
入库时间: "2026-08-12T03:23:31.257Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=latency&searchtype=all"
匹配关键词:
  - "latency"
相关厂家:
  []
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 5
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "正文为LLM用于HAProxy负载均衡的学术论文，与超节点/AI Rack/机柜级AI基础设施完全无关，不符合项目关注范围。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-12T11:23:48+08:00"
AI主题相关性: 0
AI来源权威性: 5
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 0
采集批次: "2026年8月12日2点04分20秒"
采集批次ID: "20260812-020420-289"
去重键: "https://arxiv.org/abs/2608.10532"
---

## Computer Science > Networking and Internet Architecture

## Title:Benchmarking LLM-Guided Control-Plane Policies for Backend Fault Isolation in HAProxy

Authors:[Aman Chauhan](https://arxiv.org/search/cs?searchtype=author&query=Chauhan,+A), [Vishnu Pendyala](https://arxiv.org/search/cs?searchtype=author&query=Pendyala,+V)

[View PDF](https://arxiv.org/pdf/2608.10532) [HTML (experimental)](https://arxiv.org/html/2608.10532v1)

> Abstract:Static load balancers cannot mitigate a backend that is degraded rather than down: round-robin and least-connections keep routing traffic to a server returning HTTP 500s until an operator intervenes. We ask whether a Large Language Model can replace the static routing policy itself, reading HAProxy and Prometheus telemetry every 10 seconds and isolating faulty servers through guardrailed calls to the HAProxy Data Plane API. On a reproducible benchmark with a persistent structural fault built into roughly one-third of a heterogeneous fleet, we sweep 15 open-weight models across five families (0.35B to 35B total parameters; dense, mixture-of-experts, and efficient-sparse architectures), reasoning modes, fleet scales of 3 to 9 backends, and two routing algorithms, totaling 240 runs. We find a capability threshold near 3B active parameters. Below it, LLM policies are typically unreliable and sometimes worse than no policy; above it, every model, regardless of architecture, saturates near an 88% reduction in client-perceived 5xx errors over the static baseline. The threshold is approximate: Gemma 4 E2B clears it with 2B active parameters, while the dense 3B Granite 4.0 Micro does not. The availability gain has costs. Draining concentrates load onto surviving servers, inflating tail latency 2.6 to 2.8 times, and enabling reasoning multiplies token spend roughly tenfold, overrunning the control interval and degrading effectiveness. The efficient operating point is a supra-threshold model in its cheapest non-reasoning mode, wrapped inside deterministic guardrails.

| Comments: |  |
| --- | --- |
| Subjects: | Networking and Internet Architecture (cs.NI); Machine Learning (cs.LG) |
| Cite as: | [arXiv:2608.10532](https://arxiv.org/abs/2608.10532) \[cs.NI\] |
|  | (or [arXiv:2608.10532v1](https://arxiv.org/abs/2608.10532v1) \[cs.NI\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.10532](https://doi.org/10.48550/arXiv.2608.10532) |

## Submission history

From: Aman Chauhan \[[view email](https://arxiv.org/show-email/04f2b189/2608.10532)\]  
**\[v1\]** Tue, 11 Aug 2026 06:15:20 UTC (813 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.10532) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))