---
title: Remote Direct Memory Access (RDMA)
tags: APAC HPC-AI competition
---

# RDMA(Remote Direct Memory Access)
[TOC]

---
## What is RDMA?

Remote Direct Memory Access, RDMA, enables server-to-server data movement directly between application memory without any CPU involvement. RDMA technology empowers advanced enterprise data centers and the world’s fastest supercomputers (HPC).
RDMA allows real CPU offloads and kernel bypass, and, as such, RDMA-based applications provide better utilization of network HW resources while increasing speed and lowering link latencies.

[Watch me](https://www.youtube.com/watch?v=lu78_C-9jvA)
[longer one](https://www.youtube.com/watch?v=QAIAoNheX-8)(看前面RDMA就好)
[What is RDMA article](https://www.hwchiu.com/rdma-introduction-i.html)

---
## Feature
- Zero Copy
- Kernel bypass
- No CPU involvement
- Message based transactions
- Scatter/Gather entries support

---
## Advantage
- Higher Bandwidth
- Lower Latency
- More CPU power for applications

---
## Why RDMA?

HPC 處重於高效能的演算，並透過軟體平行連結的機制串聯大量高速主機(電腦叢集 computer cluster)。可以想像成藉由網路將多個電腦串連成一台超級電腦
因此，需要RDMA的高速傳送及讀取資料，且不能影響到CPU的效能

==**Wanna go FAST? use RDMA!!**==

[RDMA + RoCE + HPC = Miracle!!](https://blogs.nvidia.com/blog/2020/04/29/what-is-rdma/)

---
## Complement

- Kernel Space
即Linux kernel的運作空間
只要是CPU可以管的硬體, kernel space的程式就可以透過 machine code來操作該硬體. 所以我們可以認為kernel space可以執行比user space更高權限的動作

- User Space
即user process的運作空間.
是比較受限的, 除了一些沒有傷害的行為之外, 基本上都不能做. 若需要在user space調用系統資源(如I/O作業), 則必須通過system call

- TCPIP

---
## Reference
- https://zh.wikipedia.org/wiki/%E8%BF%9C%E7%A8%8B%E7%9B%B4%E6%8E%A5%E5%86%85%E5%AD%98%E8%AE%BF%E9%97%AE
- https://codertw.com/%E7%A8%8B%E5%BC%8F%E8%AA%9E%E8%A8%80/587408/
- https://zhuanlan.zhihu.com/p/55142557
- https://www.itread01.com/content/1524398320.html
- https://www.youtube.com/watch?v=sPfZGHWnKNM
- https://kknews.cc/news/p6orey2.html
- https://zh.wikipedia.org/wiki/%E7%9B%B4%E6%8E%A5%E8%A8%98%E6%86%B6%E9%AB%94%E5%AD%98%E5%8F%96
相關報導:
https://kknews.cc/zh-tw/news/p6orey2.html