<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
  <title>eunomia 博客</title>
  <link>https://eunomia.dev/zh/blog/</link>
  <description>eunomia.dev 上关于 eBPF、bpftime、AI tracing 和系统研究的最新文章。</description>
  <language>zh-CN</language>
  <lastBuildDate>Fri, 06 Nov 2026 00:00:00 GMT</lastBuildDate>
  <atom:link href="https://eunomia.dev/zh/feed.xml" rel="self" type="application/rss+xml" />
  <item>
  <title>AI 写 Lean 证明的速度，已经超过 Lean 检查它们的速度</title>
  <link>https://eunomia.dev/zh/blog/2026/11/06/ai-lean-proof-checking-cost/</link>
  <guid>https://eunomia.dev/zh/blog/2026/11/06/ai-lean-proof-checking-cost/</guid>
  <pubDate>Fri, 06 Nov 2026 00:00:00 GMT</pubDate>
  <description>AI 越来越会写数学证明了。问题随之而来：它写出来的东西，我们查得过来吗？常见的回答是交给形式化验证：AI 把证明写成 Lean 代码，Lean 自动逐步核对，核对通过就说明证明成立，而机器检查又快又便宜。</description>
  <author>yunwei356@gmail.com (Yunwei 123)</author>
  <author>i@littlefisher.me (Littlefisher619)</author>
  <category>eBPF</category>
</item>
<item>
  <title>把 GPU 服务的请求循环搬进内核，能省下 CPU 吗？</title>
  <link>https://eunomia.dev/zh/blog/2026/10/17/bpf-gpu-server/</link>
  <guid>https://eunomia.dev/zh/blog/2026/10/17/bpf-gpu-server/</guid>
  <pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate>
  <description>设想一台 GPU 服务器在处理大量小请求：一个小语言模型生成 64 个 token，或者一次很短的 cuBLAS、cuFFT 调用。每个请求的 GPU 计算都很短，主机侧的开销就显得突出了。请求要经过网络协议栈，唤醒服务线程，被拷进用户态内存，排队等 GPU，最后再经过一次发送回到客户端。</description>
  <author>i@littlefisher.me (yunwei37)</author>
  <category>eBPF</category>
</item>
<item>
  <title>When Does CPU Scheduling Help AI Inference?</title>
  <link>https://eunomia.dev/zh/blog/2026/10/17/gthulhu-vllm-cpu-pressure/</link>
  <guid>https://eunomia.dev/zh/blog/2026/10/17/gthulhu-vllm-cpu-pressure/</guid>
  <pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate>
  <description>LLM inference is usually discussed as a GPU problem. Matrix multiplication, attention, KV-cache access, and model execution dominate the compute budget, and that framing is mostly right. But a production inference</description>
  <author>i@littlefisher.me (Littlefisher619)</author>
  <author>iancodinghtml@gmail.com (Ian Chen)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>语言模型能在 Linux eBPF 里运行吗？</title>
  <link>https://eunomia.dev/zh/blog/2026/09/24/qwen3-ebpf-in-kernel-inference/</link>
  <guid>https://eunomia.dev/zh/blog/2026/09/24/qwen3-ebpf-in-kernel-inference/</guid>
  <pubDate>Thu, 24 Sep 2026 00:00:00 GMT</pubDate>
  <description>我们通常用 eBPF 观察或影响别的程序：跟踪系统调用、过滤网络包，或者调整调度策略。qwen3-ebpf 提了一个不同的问题：通过内核验证的 BPF 程序，能否亲自完成真实语言模型的数值计算，而不只是观察模型服务？</description>
  <author>i@littlefisher.me (Littlefisher619)</author>
  <category>eBPF</category>
</item>
<item>
  <title>为什么 eBPF verifier 报错难修：诊断鸿沟</title>
  <link>https://eunomia.dev/zh/blog/2026/07/25/ebpf-verifier-errors-bpfix/</link>
  <guid>https://eunomia.dev/zh/blog/2026/07/25/ebpf-verifier-errors-bpfix/</guid>
  <pubDate>Sat, 25 Jul 2026 00:00:00 GMT</pubDate>
  <description>当开发者将 eBPF 程序加载到 Linux 内核时，verifier 必须在字节码执行之前证明程序安全。verifier 遍历程序的每一条可能执行路径，逐指令跟踪它对每个寄存器和栈槽的了解。如果发现某条指令无法证明安全，例如通过未验证的指针访问内存、读取超出数据包边界或存在无界循环，它就会拒绝程序并打印报错。</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <category>eBPF</category>
</item>
<item>
  <title>从 Token 到可验证工作：Agent Infra 正在改写它的计量单位</title>
  <link>https://eunomia.dev/zh/blog/2026/07/20/agent-work-unit/</link>
  <guid>https://eunomia.dev/zh/blog/2026/07/20/agent-work-unit/</guid>
  <pubDate>Mon, 20 Jul 2026 00:00:00 GMT</pubDate>
  <description>一次模型调用容易记账：哪个模型、多少 token、多长延迟、有没有报错。Agent 打破了这种简洁。一次用户请求现在可能跨越几十次调用、多个工具、人工审批、重试、外部系统的状态变化。调用仍然可观测，但工作有没有真的完成，是另一个问题。</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <category>daily-analysis</category>
  <category>research</category>
  <category>AI Agent</category>
  <category>Agent Infrastructure</category>
  <category>Observability</category>
</item>
<item>
  <title>实证研究：AI Agent 规则需要上下文与分层强制执行</title>
  <link>https://eunomia.dev/zh/blog/2026/07/15/ebpf-ai-agent-policy-enforcement/</link>
  <guid>https://eunomia.dev/zh/blog/2026/07/15/ebpf-ai-agent-policy-enforcement/</guid>
  <pubDate>Wed, 15 Jul 2026 00:00:00 GMT</pubDate>
  <description>“提交前必须跑完整测试”这种规则看起来很简单，直到 AI Agent 在上次测试后又改了源码，然后继续执行 git commit。内核只看见普通进程在写提交对象，负责管理 Agent 工具调用与会话状态的应用层也只看见又一次工具调用，但真正决定这次提交能不能放行的，是哪次测试结果仍然有效、哪次编辑让它失效，以及规则在当前任务里该如何实例化。</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <author>i@littlefisher.me (LinuxDev9002)</author>
  <author>1067852565@qq.com (云微)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>让 AI Agent 调优 Linux 调度器：SchedCP 与 sched-agent 的设计与评测</title>
  <link>https://eunomia.dev/zh/blog/2026/07/10/schedcp-agentic-linux-scheduler/</link>
  <guid>https://eunomia.dev/zh/blog/2026/07/10/schedcp-agentic-linux-scheduler/</guid>
  <pubDate>Fri, 10 Jul 2026 00:00:00 GMT</pubDate>
  <description>Linux 调度器根据可运行任务、唤醒事件、优先级和累积的 CPU 时间做出决策，应用却用另一套指标评价结果：编译内核时关心最后一个编译进程何时结束（makespan），在线服务盯着 P99 延迟和吞吐量，一组并行的批处理任务可能被其中最慢的那个拖住（长尾分布）。内核提供机制和测量手段，工作负载决定怎样才算“更好”。这种错位（论文称之为“语义鸿沟”）意味着默认的 EEVDF 调度器在面对需要不同权衡的工作负载时只能用一套通用策略应对。</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <author>1067852565@qq.com (云微)</author>
  <author>i@littlefisher.me (LinuxDev9002)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>AI Agent Trace 的语义 Flamegraph</title>
  <link>https://eunomia.dev/zh/blog/2026/06/24/agentpprof-semantic-flamegraph/</link>
  <guid>https://eunomia.dev/zh/blog/2026/06/24/agentpprof-semantic-flamegraph/</guid>
  <pubDate>Wed, 24 Jun 2026 00:00:00 GMT</pubDate>
  <description>AI Agent 跑完一个月，你知道 token 总量却不知道分布。代码审查花了多少、调试花了多少？工具重试是否显著、continuation prompt 是否占主导？这些问题关乎工作流优化，但传统可观测性工具无法直接回答。</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <author>1067852565@qq.com (云微)</author>
  <author>i@littlefisher.me (LinuxDev9002)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>实证研究：AI Agent 规则需要上下文与分层强制执行</title>
  <link>https://eunomia.dev/zh/blog/2026/05/31/actplane-pushing-agent-harness-enforcement-down-to-kernel-ebpf/</link>
  <guid>https://eunomia.dev/zh/blog/2026/05/31/actplane-pushing-agent-harness-enforcement-down-to-kernel-ebpf/</guid>
  <pubDate>Sun, 31 May 2026 00:00:00 GMT</pubDate>
  <description>“提交前必须跑测试”这条规则看起来很简单，直到 AI Agent 在上次测试后又改了源码，然后直接执行 git commit。内核只看见普通进程在写提交对象，harness 也只看见又一次工具调用，但真正决定这次提交能否放行的，是哪次测试结果仍然有效、哪次编辑让它失效，以及规则在当前任务里该如何实例化。</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <author>i@littlefisher.me (LinuxDev9002)</author>
  <author>1067852565@qq.com (云微)</author>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>基于 eBPF 的不透明 AI Agent 运行时可观测与执行控制：超越沙箱与审批</title>
  <link>https://eunomia.dev/zh/blog/2026/05/25/runtime-security-for-ai-agents/</link>
  <guid>https://eunomia.dev/zh/blog/2026/05/25/runtime-security-for-ai-agents/</guid>
  <pubDate>Mon, 25 May 2026 00:00:00 GMT</pubDate>
  <description>AI 编程 Agent 如今可以连续运行数小时，端到端地完成整个功能开发，优化生产级 GPU 内核，并自主合并数千个 Pull Request。然而，大多数 Agent 安全仍然依赖 人工审批（human-in-the-loop），而 Anthropic 自己的数据显示用户在 93% 的权限 提示上直接点了同意，缺乏实质性审查。结果可以预见：产品纷纷添加绕过模式， 用户关闭权限门控，65% 的企业报告了 Agent 安全事件。</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>CPU 噪声会拖慢 GPU 推理吗：用 eBPF 定量测量调度器与 IRQ 影响</title>
  <link>https://eunomia.dev/zh/blog/2026/05/24/measuring-cpu-noise-in-gpu-inference-with-ebpf/</link>
  <guid>https://eunomia.dev/zh/blog/2026/05/24/measuring-cpu-noise-in-gpu-inference-with-ebpf/</guid>
  <pubDate>Sun, 24 May 2026 00:00:00 GMT</pubDate>
  <description>GPU 推理看起来像是 GPU 问题，但 CPU 仍然在关键路径上。CPU 负责准备输入、发起 CUDA kernel launch、管理同步和运行时调用，同时还要和系统任务、中断、其他租户共享核心。如果 CPU 端 launch 路径被打断，GPU 可能会在有算力的情况下等不到新任务。</description>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <category>eBPF</category>
</item>
<item>
  <title>ACRFence：防止 AI Agent 检查点恢复中的语义回滚攻击</title>
  <link>https://eunomia.dev/zh/blog/2026/05/21/acrfence-preventing-semantic-rollback-attacks-in-agent-checkpoint-restore/</link>
  <guid>https://eunomia.dev/zh/blog/2026/05/21/acrfence-preventing-semantic-rollback-attacks-in-agent-checkpoint-restore/</guid>
  <pubDate>Thu, 21 May 2026 00:00:00 GMT</pubDate>
  <description>AI Agent 框架正在把 checkpoint/restore、time travel、rewind 这类能力带进日常开发体验：Agent 做错了，可以回到某个检查点；用户想试另一条路径，可以从旧状态分叉继续跑。这对调试和人机协作非常有用，但当 Agent 已经调用过外部工具时，事情会变得危险。</description>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>AgentCgroup：当 AI Agent 遇到操作系统资源</title>
  <link>https://eunomia.dev/zh/blog/2026/02/17/agentcgroup-what-happens-when-ai-coding-agents-meet-os-resources/</link>
  <guid>https://eunomia.dev/zh/blog/2026/02/17/agentcgroup-what-happens-when-ai-coding-agents-meet-os-resources/</guid>
  <pubDate>Tue, 17 Feb 2026 00:00:00 GMT</pubDate>
  <description>一个 AI Agent 安静地读了几分钟代码，改完文件后启动 pytest。测试进程加载依赖时，内存可能在一秒内增加数百 MB，命令退出后又迅速回落。容器级控制器只看到整个工作负载越过阈值，却分不清这次突发来自短命的工具进程，而长时间运行的 Agent 运行时还保存着对话、诊断过程和未完成的修改。</description>
  <author>i@littlefisher.me (Littlefisher619)</author>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <author>1067852565@qq.com (云微)</author>
  <author>i@littlefisher.me (LinuxDev9002)</author>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>Reverse Engineering Claude Code&#39;s SSL Traffic with eBPF</title>
  <link>https://eunomia.dev/zh/blog/2026/02/13/reverse-engineering-claude-codes-ssl-traffic-with-ebpf/</link>
  <guid>https://eunomia.dev/zh/blog/2026/02/13/reverse-engineering-claude-codes-ssl-traffic-with-ebpf/</guid>
  <pubDate>Fri, 13 Feb 2026 00:00:00 GMT</pubDate>
  <description>Ever wondered what your AI coding agent is actually sending over the network? As part of our work on AgentSight — an eBPF-powered observability tool that monitors AI agents at system boundaries without modifying their</description>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <category>eBPF</category>
</item>
<item>
  <title>A Taxonomy of GPU Bugs: 19 Defect Classes for CUDA Verification</title>
  <link>https://eunomia.dev/zh/blog/2026/01/29/a-taxonomy-of-gpu-bugs-19-defect-classes-for-cuda-verification/</link>
  <guid>https://eunomia.dev/zh/blog/2026/01/29/a-taxonomy-of-gpu-bugs-19-defect-classes-for-cuda-verification/</guid>
  <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
  <description>GPU programming introduces a distinct class of correctness and performance challenges that differ fundamentally from traditional CPU-based systems.</description>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <author>1067852565@qq.com (云微)</author>
  <category>eBPF</category>
</item>
<item>
  <title>智能体系统架构：隔离、集成与治理的综合调研</title>
  <link>https://eunomia.dev/zh/blog/2026/01/11/architectures-for-agent-systems-a-survey-of-isolation-integration-and-governance/</link>
  <guid>https://eunomia.dev/zh/blog/2026/01/11/architectures-for-agent-systems-a-survey-of-isolation-integration-and-governance/</guid>
  <pubDate>Sun, 11 Jan 2026 00:00:00 GMT</pubDate>
  <description>基于大语言模型（LLM）的智能体系统，即能够利用LLM自主规划并借助外部工具执行多步任务的软件，正迅速从概念验证阶段迈向企业级规模化部署。这些智能体有望将编码、IT运维、数据分析等工作自动化，然而，将其部署于生产环境也带来了在安全性、可靠性及系统集成方面的一系列新挑战。过去半年，业界社区已在几项关键策略上形成共识：为执行不可信操作提供强有力的隔离保障，为工具集成建立标准化协议，并构建能确保智能体行为符合企业策略的治理框架。本调研旨在系统</description>
  <author>i@littlefisher.me (Littlefisher)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>GPU可观测性差距：为什么我们需要卸载到GPU设备上的eBPF</title>
  <link>https://eunomia.dev/zh/blog/2025/10/14/the-gpu-observability-gap-why-we-need-ebpf-on-gpu-devices/</link>
  <guid>https://eunomia.dev/zh/blog/2025/10/14/the-gpu-observability-gap-why-we-need-ebpf-on-gpu-devices/</guid>
  <pubDate>Tue, 14 Oct 2025 00:00:00 GMT</pubDate>
  <description>eBPF 作为一种在内核态提供可编程能力的革命性技术，已经在 CPU 的可观测性、网络和安全领域取得了巨大成功。然而，对于日益重要的 GPU 计算领域，我们同样需要一种灵活、高效的观测手段。目前，大多数 GPU 性能分析工具都局限于在 CPU 侧通过驱动/用户态 API 或厂商的性能分析接口（CUPTI）进行观测，难以深入了解 GPU 内部的执行细节。为了解决这一问题，bpftime 通过其 CUDA/SYCL 附加实现提供 GPU</description>
  <author>i@littlefisher.me (Littlefisher)</author>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <category>eBPF</category>
</item>
<item>
  <title>NVIDIA Open GPU Kernel Modules Comprehensive Source Code Analysis</title>
  <link>https://eunomia.dev/zh/blog/2025/10/14/nvidia-open-gpu-kernel-modules-comprehensive-source-code-analysis/</link>
  <guid>https://eunomia.dev/zh/blog/2025/10/14/nvidia-open-gpu-kernel-modules-comprehensive-source-code-analysis/</guid>
  <pubDate>Tue, 14 Oct 2025 00:00:00 GMT</pubDate>
  <description>In May 2022, NVIDIA made a decision that would fundamentally alter the landscape of GPU computing on Linux: they open-sourced the kernel-mode components of their GPU driver.</description>
  <author>1067852565@qq.com (云微)</author>
  <author>linuxdev8883@example.com (LinuxDev9002)</author>
  <author>i@littlefisher.me (Littlefisher)</author>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <category>eBPF</category>
</item>
<item>
  <title>Understanding iaprof: A Deep Dive into AI/GPU Flame Graph Profiling</title>
  <link>https://eunomia.dev/zh/blog/2025/10/11/understanding-iaprof-a-deep-dive-into-ai-gpu-flame-graph-profiling/</link>
  <guid>https://eunomia.dev/zh/blog/2025/10/11/understanding-iaprof-a-deep-dive-into-ai-gpu-flame-graph-profiling/</guid>
  <pubDate>Sat, 11 Oct 2025 00:00:00 GMT</pubDate>
  <description>An exploration of Intel&#39;s innovative profiling tool that bridges the gap between CPU and GPU execution</description>
  <author>yuxi4096@gmail.com (yuxi4096)</author>
  <author>1067852565@qq.com (云微)</author>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <category>eBPF</category>
</item>
<item>
  <title>AgentSight: 让 AI Agent 的一举一动都在掌控之中，基于 eBPF 的系统级可观测性方案</title>
  <link>https://eunomia.dev/zh/blog/2025/08/26/agentsight-keeping-your-ai-agents-under-control-with-ebpf-powered-system-observability/</link>
  <guid>https://eunomia.dev/zh/blog/2025/08/26/agentsight-keeping-your-ai-agents-under-control-with-ebpf-powered-system-observability/</guid>
  <pubDate>Tue, 26 Aug 2025 00:00:00 GMT</pubDate>
  <description>想象一下，当你的 AI Agent 正在自主编写代码、执行命令时，你却不知道它究竟在做什么，或许有点令人不安？随着 Claude Code、Gemini-cli 等 LLM 智能体工具的普及，我们正在将越来越多的控制权交给 AI。但这里有个棘手的问题：这些 AI 智能体与传统软件截然不同，现有的监控工具要么只能看到它们的&quot;想法&quot;（LLM</description>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>Profiling and Tracing Tools Across System Layers and Architectures</title>
  <link>https://eunomia.dev/zh/blog/2025/08/20/profiling-and-tracing-tools-across-system-layers-and-architectures/</link>
  <guid>https://eunomia.dev/zh/blog/2025/08/20/profiling-and-tracing-tools-across-system-layers-and-architectures/</guid>
  <pubDate>Wed, 20 Aug 2025 00:00:00 GMT</pubDate>
  <description>Profiling and tracing are complementary techniques for analyzing software performance and behavior.</description>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <category>eBPF</category>
</item>
<item>
  <title>The Modern Memory Testing Arsenal -- A Complete Guide to Benchmarking Tools for Next-Gen Memory Systems</title>
  <link>https://eunomia.dev/zh/blog/2025/06/21/the-modern-memory-testing-arsenal-a-complete-guide-to-benchmarking-tools-for-next-gen-memory-systems/</link>
  <guid>https://eunomia.dev/zh/blog/2025/06/21/the-modern-memory-testing-arsenal-a-complete-guide-to-benchmarking-tools-for-next-gen-memory-systems/</guid>
  <pubDate>Sat, 21 Jun 2025 00:00:00 GMT</pubDate>
  <description>Memory systems are evolving rapidly. From traditional DDR DRAM to high-bandwidth memory (HBM), persistent memory (PMEM), and the emerging Compute Express Link (CXL) technology, today&#39;s systems feature complex</description>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <category>eBPF</category>
</item>
<item>
  <title>系统会议中的可观测性、性能分析和调试（2015–2025）</title>
  <link>https://eunomia.dev/zh/blog/2025/06/21/observability-profiling-and-debugging-in-systems-conference-2015-2025/</link>
  <guid>https://eunomia.dev/zh/blog/2025/06/21/observability-profiling-and-debugging-in-systems-conference-2015-2025/</guid>
  <pubDate>Sat, 21 Jun 2025 00:00:00 GMT</pubDate>
  <description>本综述回顾了十多年来（2015–2025）计算机系统可观测性、性能分析和调试技术的研究，重点关注OSDI、SOSP和EuroSys的主会论文。我们涵盖了100多篇论文，涉及动态跟踪框架、日志记录和监控基础设施、性能异常检测、根因分析和系统可见性机制。我们识别了所解决的核心问题（从分布式请求跟踪到配置或并发错误检测）、采用的技术（动态插桩、静态分析、原位日志记录、分布式监控、ML辅助分析）、目标领域（操作系统内核、云和分布式系统、移动/物</description>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <category>eBPF</category>
</item>
<item>
  <title>Checkpoint/Restore Systems: Evolution, Techniques, and Applications in AI Agents</title>
  <link>https://eunomia.dev/zh/blog/2025/05/11/checkpoint-restore-systems-evolution-techniques-and-applications-in-ai-agents/</link>
  <guid>https://eunomia.dev/zh/blog/2025/05/11/checkpoint-restore-systems-evolution-techniques-and-applications-in-ai-agents/</guid>
  <pubDate>Sun, 11 May 2025 00:00:00 GMT</pubDate>
  <description>Checkpoint/restore (C/R) technology – the ability to save a running program’s state to persistent storage and later resume execution from that point – has long been a cornerstone of fault tolerance and process</description>
  <author>yt.xyxx@gmail.com (officeyutong)</author>
  <category>AI agent observability</category>
</item>
<item>
  <title>ASPLOS 2025: Paper Summaries and Insights</title>
  <link>https://eunomia.dev/zh/blog/2025/04/25/asplos-2025-paper-summaries-and-insights/</link>
  <guid>https://eunomia.dev/zh/blog/2025/04/25/asplos-2025-paper-summaries-and-insights/</guid>
  <pubDate>Fri, 25 Apr 2025 00:00:00 GMT</pubDate>
  <description>The Association for Computing Machinery&#39;s Architectural Support for Programming Languages and Operating Systems (ASPLOS) conference is a premier venue where researchers present cutting-edge work spanning computer</description>
  <author>yt.xyxx@gmail.com (officeyutong)</author>
  <category>eBPF</category>
</item>
<item>
  <title>EuroSys 2025 Paper Summaries and Analysis</title>
  <link>https://eunomia.dev/zh/blog/2025/04/25/eurosys-2025-paper-summaries-and-analysis/</link>
  <guid>https://eunomia.dev/zh/blog/2025/04/25/eurosys-2025-paper-summaries-and-analysis/</guid>
  <pubDate>Fri, 25 Apr 2025 00:00:00 GMT</pubDate>
  <description>This post offers a detailed examination of papers accepted at EuroSys 2025, one of the premier conferences in computer systems research.</description>
  <author>yt.xyxx@gmail.com (officeyutong)</author>
  <category>eBPF</category>
</item>
<item>
  <title>深入GPU性能分析工具：现代加速器追踪工具的实现详解</title>
  <link>https://eunomia.dev/zh/blog/2025/04/21/gpu-profiling-under-the-hood-an-implementation-focused-survey-of-modern-accelerator-tracing-tools/</link>
  <guid>https://eunomia.dev/zh/blog/2025/04/21/gpu-profiling-under-the-hood-an-implementation-focused-survey-of-modern-accelerator-tracing-tools/</guid>
  <pubDate>Mon, 21 Apr 2025 00:00:00 GMT</pubDate>
  <description>对异构加速器（GPU、DPU和APU）进行性能分析和追踪对于优化现代系统的性能至关重要。本调研提供了一个深入的实现导向的综述，介绍了当前最先进的工具如何捕获低级执行细节。我们面向两类受众：(1) 工具开发者，他们希望了解现有性能分析工具的构建方式——包括它们钩住的接口以及如何追踪CPU和加速器；(2)</description>
  <author>yunwei356@gmail.com (yunwei37)</author>
  <author>yt.xyxx@gmail.com (officeyutong)</author>
  <category>eBPF</category>
</item>
<item>
  <title>加速器工具箱：GPU和其他协处理器的性能分析和追踪详解</title>
  <link>https://eunomia.dev/zh/blog/2025/04/11/the-accelerator-toolkit-a-review-of-profiling-and-tracing-for-gpus-and-other-co-processor/</link>
  <guid>https://eunomia.dev/zh/blog/2025/04/11/the-accelerator-toolkit-a-review-of-profiling-and-tracing-for-gpus-and-other-co-processor/</guid>
  <pubDate>Fri, 11 Apr 2025 00:00:00 GMT</pubDate>
  <description>现代计算越来越依赖专用加速器——尤其是GPU、DPU和APU——来处理各种工作负载。图形处理单元(GPU)是一种大规模并行处理器，最初用于图形处理，现在在通用计算(HPC)和AI中必不可少。数据处理单元(DPU)是一种较新类型的可编程处理器，结合了CPU核心和高性能网络/存储引擎。DPU从CPU卸载网络、安全和存储任务，被认为是继CPU和GPU之后计算的&quot;第三支柱&quot;。同时，加速处理单元(APU)在一个芯片上集成了CPU和GPU组件——这</description>
  <author>yt.xyxx@gmail.com (officeyutong)</author>
  <category>eBPF</category>
</item>
<item>
  <title>OS-Level Challenges in LLM Inference and Optimizations</title>
  <link>https://eunomia.dev/zh/blog/2025/02/18/os-level-challenges-in-llm-inference-and-optimizations/</link>
  <guid>https://eunomia.dev/zh/blog/2025/02/18/os-level-challenges-in-llm-inference-and-optimizations/</guid>
  <pubDate>Tue, 18 Feb 2025 00:00:00 GMT</pubDate>
  <description>Large Language Model (LLM) inference pushes computing systems to their limits, not only in raw compute but also in how the operating system (OS) manages resources.</description>
  <author>yusen356@gmail.com (weekendfish)</author>
  <author>1067852565@qq.com (云微)</author>
  <category>AI agent observability</category>
</item>
</channel>
</rss>