Architecture

• Tutti System Architecture
  • 1. Overview
  • 2. Layered architecture
  • 3. IO walkthrough
  • 4. Deployment topology
  • 5. Source map
• Key Designs — Performance & Low Overhead
  • 1. GPU io_uring — shared submission state, one hand-off per batch, async harvest
  • 2. Register-time precomputation — the hot path is table lookup, not arithmetic
  • 3. Two-tier caches sized for LMCache-scale file counts
  • 4. A batched, event-frugal control plane
  • 5. Multi-device striping
• snvme 设计依据：为什么需要一个定制的 NVMe 驱动
  • 1. snvme 做的三件不可替代的事
  • 2. 现有内核基础设施逐项对照
  • 3. 根本阻塞点（不随内核版本变化）
  • 4. RDMA 世界的参照：用户态直敲 doorbell 是什么
  • 5. RDMA ↔ snvme 逐项对照
  • 6. 替代路径：上游化