https://yangyang233333.github.io/tags/storageruntime/ Recent content in StorageRuntime on Hellokitty's Blog Hugo zh-CN Thu, 03 Sep 2026 17:35:00 +0800 https://yangyang233333.github.io/posts/tutti-runtime-source-reading/ Thu, 03 Sep 2026 17:35:00 +0800 https://yangyang233333.github.io/posts/tutti-runtime-source-reading/ <p>Tutti 的论文强调 GPU 主导 I/O，但应用首先接触的不是 NVMe doorbell，而是一套稳定的 <code>StorageRuntime</code> API。它负责把 URI、文件 extent、GPU 内存、后端实例和异步请求组装起来，让上层无需知道底层究竟是单盘、条带化 NVMe，还是未来的其他传输路径。</p> <p>本文阅读 Tutti v0.1.1 的运行时源码与设计文档，聚焦五个问题：公开 API 长什么样、<code>open</code> 如何解析对象、<code>register_memory</code> 为什么独立存在、<code>submit/wait</code> 如何路由，以及 vLLM connector 如何接入。</p> <p>项目源码：<a href="https://github.com/xPU-IO/Tutti">xPU-IO/Tutti</a>。</p> <h2 id="一先看整体分层">一、先看整体分层</h2> <p>当前代码把系统拆成以下层次：</p> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;-webkit-text-size-adjust:none;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>Application / vLLM Connector </span></span><span style="display:flex;"><span> │ </span></span><span style="display:flex;"><span> ▼ </span></span><span style="display:flex;"><span> StorageRuntime </span></span><span style="display:flex;"><span> open · register · submit · wait </span></span><span style="display:flex;"><span> │ </span></span><span style="display:flex;"><span> ┌──────┴──────┐ </span></span><span style="display:flex;"><span> ▼ ▼ </span></span><span style="display:flex;"><span> Resolver DataPath </span></span><span style="display:flex;"><span> URI → target IO → backend </span></span><span style="display:flex;"><span> │ │ </span></span><span style="display:flex;"><span> └──── Binding ┘ </span></span><span style="display:flex;"><span> │ </span></span><span style="display:flex;"><span> local / striped NVMe </span></span></code></pre></div><p><code>StorageRuntime</code> 是唯一稳定的应用接口；Resolver、Binding 和 DataPath 是内部 SPI。这样做避免上层依赖特定 NVMe 驱动或 GPU 厂商实现。</p>