跳转至

面向高扇出智能体沙箱的内存压缩技术

Memory Compression for High-Fanout Agent Sandboxes

arXiv ID: arXiv:2609.11294 [cs.AI]
Subjects: Artificial Intelligence (cs.AI); Operating Systems (cs.OS)
Submission Date: 10 September 2026
Authors: Mengming Li, Ceyu Xu, Qijun Zhang, Jiangnan Yu, Xiangfeng Sun, Haohui Mai, Zhiyao Xie

文章背景与核心概要

随着高扇出 (High-Fanout) AI 智能体应用的大规模普及,单个复合任务往往需要同时并发生成数十上百个沙箱容器来进行代码测试与环境交互,这给宿主机系统带来了极其沉重的物理内存瓶颈。然而,这些沙箱并非完全独立,它们源自相同的初始模板并执行高度相似的代码轨迹,存在海量的跨沙箱与相对模板的内存冗余。传统操作系统的通用内存压缩机制无法识别这种智能体执行特征,面临压缩方式僵化、压缩范围保守以及触发时机被动的三大局限。为此,本文提出了专为 AI 智能体沙箱量身定制的内存压缩系统 AgentZip,通过挖掘模板相对差异与跨沙箱冗余、以恢复期预取替代压缩期保守筛选、并将重度压缩操作对齐到模型等待空闲期。在涵盖大语言模型训练与推理的多样化工作负载中,AgentZip 实现了最高达 8.7 倍的沙箱内存缩减,同时将激进压缩带来的执行延迟从 3.1 倍大幅降低至 1.40 倍,为大规模智能体并发执行提供了坚实高效的底层系统支撑。


核心概要

Summary

高扇出 AI 智能体 (AI Agent) 工作负载由于从共享模板中衍生出大量并发沙箱会话,从而造成了巨大的系统内存瓶颈。尽管传统的内存压缩机制由于压缩策略、作用范围和压缩时机的不匹配而难以应对这种场景,但本文提出了 AgentZip。AgentZip 专为 AI 智能体沙箱打造,深度利用了相对模板与跨沙箱的内存冗余,通过在恢复时进行页面预取来优化页面选择,并将耗时的压缩阶段与大语言模型 (Large Language Model, LLM) 的等待时间对齐。这不仅显著削减了内存占用,同时将任务执行减速控制在极低水平。

High-fanout AI agent workloads create a massive memory bottleneck by spawning numerous concurrent sandbox sessions from shared templates. While traditional memory compression struggles with this due to mismatched compression strategies, scopes, and timings, this paper introduces AgentZip. Specifically designed for AI-agent sandboxes, AgentZip leverages template-relative and cross-sandbox redundancies, optimizes page selection via restore-time prefetching, and aligns compression phases with LLM waiting times. This dramatically reduces memory usage while minimizing execution slowdowns.


论文元数据

Paper Metadata

  • arXiv ID: arXiv:2609.11294 [cs.AI]
  • 学科领域: 人工智能 (cs.AI);操作系统 (cs.OS)
  • 提交日期: 2026 年 9 月 10 日
  • 作者列表:
  • Mengming Li
  • Ceyu Xu
  • Qijun Zhang
  • Jiangnan Yu
  • Xiangfeng Sun
  • Haohui Mai
  • Zhiyao Xie
  • 全文获取:
  • 查看 PDF
  • 网页版本
  • TeX 源码
  • arXiv ID: arXiv:2609.11294 [cs.AI]
  • Subjects: Artificial Intelligence (cs.AI); Operating Systems (cs.OS)
  • Submission Date: 10 September 2026
  • Authors:
  • Mengming Li
  • Ceyu Xu
  • Qijun Zhang
  • Jiangnan Yu
  • Xiangfeng Sun
  • Haohui Mai
  • Zhiyao Xie
  • Full-Text Access:
  • View PDF
  • HTML Version
  • TeX Source

论文摘要

Abstract

高扇出智能体工作负载正在引发日益严峻的内存瓶颈,因为单个复杂任务就可能催生出众多并发运行的沙箱会话。然而,这些沙箱绝非相互独立:它们不仅派生自同一个基础模板,而且执行着高度相关的运行轨迹,从而暴露出可观的相对模板冗余与跨沙箱内存冗余。

High-fanout agent workloads create a growing memory bottleneck because a single task may spawn many concurrent sandbox sessions. Yet these sandboxes are far from independent: they originate from a shared template and execute related trajectories, exposing substantial template-relative and cross-sandbox memory redundancy.

传统的内存压缩技术在三个根本维度上与此类场景严重错配: 1. 如何压缩: 无法有效挖掘非完全相同沙箱内存页之间的相似性。 2. 压缩什么: 只能通过保守地筛选页面来被动控制缺页异常开销。 3. 何时压缩: 压缩要么在遭遇内存严重压力时被动触发,要么在对智能体执行阶段毫无感知的情况下盲目运行。

Conventional memory compression is poorly matched to this setting in three fundamental dimensions: 1. How to compress: They fail to exploit similarity across non-identical sandbox pages. 2. What to compress: They control page-fault overhead through conservative page selection. 3. When to compress: Compression is either triggered reactively by memory pressure or performed without awareness of agent execution phases.

为了攻克上述难题,作者团队推出了 AgentZip——首个专为 AI 智能体沙箱量身打造的内存压缩系统: * 深度挖掘冗余: 引入全新机制,充分利用相对模板冗余与跨沙箱数据冗余。 * 扩大范围与主动预取: 将压缩作用域拓展到任何具有压缩增益的页面,将开销控制的重心从压缩时的保守挑选转移至恢复时的页面预取。 * 执行阶段感知调度: 将计算密集型的压缩任务精准对齐到大模型的等待空闲期,避免对前台工具的高效执行产生干扰。

To resolve these challenges, the authors present AgentZip, the first memory compression system designed specifically for AI-agent sandboxes: * Advanced Redundancy Exploitation: Introduces mechanisms that leverage both template-relative and cross-sandbox redundancies. * Expanded Scope & Prefetching: Broadens the compression scope to any page with a profitable representation, shifting overhead control from compression-time page selection to restore-time prefetching. * Phase-Aware Scheduling: Aligns expensive compression tasks with LLM waiting periods to avoid interfering with foreground tool execution.

实验结果

Results

在涵盖大语言模型训练与推理的多样化工作负载中,AgentZip 将沙箱独占内存削减了高达 8.7 倍(相比之下标准 Linux 配置仅为 2.1 倍)。此外,恢复期预取与智能体执行阶段感知调度机制成功将激进压缩导致的性能减速从高达 3.1 倍大幅压降至 1.40 倍,且几乎完整保留了全部的内存节约收益。

Across LLM training and inference workloads, AgentZip reduces sandbox-owned memory by up to 8.7x (compared to 2.1x for standard Linux configurations). Furthermore, restore prefetching and agent-execution-aware scheduling successfully reduce the slowdown of aggressive compression from as high as 3.1x down to 1.40x while retaining nearly all of its memory-saving benefits.


授权与相关资源

License & Resources