Try   HackMD
tags: Session Three

In-network congestion : Adaptive Routing(應凱軒)

MPIGraph static route and adaptive route

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

Static Route (靜態路由)

  • 靜態路由必須由管理人員手動輸入，好處是速度很快，不需要經過學習。但是缺點就是網路拓撲若有任何的改變，管理人員必須更新資料到Router設備之中，比較麻煩，也耗費人力，必須有個很好的維護能力才行。這種方式比較適合幾乎不會有變動的網路拓撲。

Adaptive Route = Dynamic Route(自適應路由、動態路由)

  • 動態路由就不需要手動輸入，比較屬於Worry-Free的類型，一切的工作都交給Router設備之間去協調，互相交換並學習。管理人員只要做Routing Protocol的設定即可，但是這種方式比較耗費系統資源，速度也稍微慢

  • 有關這張圖的補充及出處
    https://www.hpcwire.com/2020/03/16/gpcnet-or-gpcnot/

small message size and large message size

  • small message size
    • Message Rate
    • Non-contiguous data access
  • 封包傳輸速率是主要的問題
  • 封包聚合 將小封包聚合在一起，進而漸少封包傳送的次數和封包多餘的標頭。
  • large message size
    • congestion

Nonblocking Collective Operations

  • 非阻塞系統是指在不能立刻得到調用结果之前，函式不會阻塞當前線程，而會立刻返回的I/O系统。調用非阻塞I/O跟阻塞I/O的差别為調用之后立即返回，返回后，CPU的時間片可以用来處理其他事務，此時性能是提升的。
  • in order(MPI model) 所有進程必須以相同的順序調用每個通信器的集合操作
  • tagged operation (AI/DL model) 啟動集合操作無順序要求（需要專用資源）
Last changed by 
 
HPC-AI-group
0
328

Read more

HPC Sessions

RDMA Fundation RDMA簡介 Zero Copy(林俊諺) Transport offload/Introduction to verb/Verbs object(紀見如) data-path flow/memory management/verbs in short (應凱軒) RDMA send/receive和write/read(葉鎮邦） Atomic/Memory Registration/Send Request (羅君瑜) Receive Request(吳侑珉) RDMA Programming - 1

Aug 29, 2021
Transport offload/Introduction to verb/Verbs object(紀見如)

Transport offload Enable data transfer between applicatuin on hosts 1.Reliability 可靠性 補充 : 2.Connectivity 連接性 3.Message

Aug 23, 2021
UCX Api Overview

UCX Example(1) - Init What is context? UCP application context (or just a context) is an opaque handle that ==holds a UCP communication instance’s global information==. It represents a single UCP communication instance. The communication instance could be an OS process (an application) that uses UCP library. This global information includes ==communication resources, endpoints, memory, temporary file storage, and other communication information directly associated with a specific UCP instance==. The context also acts as an isolation mechanism, allowing resources associated with the context to manage multiple concurrent communication instances. For example, users using both MPI and OpenSHMEM sessions simultaneously can isolate their communication by allocating and using separate contexts for each of them. Alternatively, users can share the communication resources (memory, network resource context, etc.) between them by using the same application context. A message sent or a RMA operation performed in one application context cannot be received in any other application context. (轉自UCX Docs) What is worker? UCP worker is an ==opaque object representing the communication context==. The worker represents an instance of a local communication resource and the progress engine associated with it. The ==progress engine is a construct that is responsible for asynchronous== and independent progress of communication directives. The progress engine could be implemented in hardware or software. ==The worker object abstracts an instance of network resources such as a host channel adapter port, network interface, or multiple resources such as multiple network interfaces or communication ports==. It could also represent virtual communication resources that are defined across multiple devices. Although the worker can represent multiple network resources, it is associated with a single UCX application context. All communication functions require a context to perform the operation on the dedicated hardware resource(s) and an endpoint to address the destination.

Aug 22, 2021
ENternet vs. InfiniBand

網路協議 廣播風暴 在這個網路拓樸下，如果沒有開啟 STP 會發生什麼事呢 首先，當 PC A 發出了一個 L2 的廣播封包（像是之前介紹的 ARP 請求封包，就是廣播形式的） PC A 將會發送至 Switch 2（他也只有連上 Switch 2，也沒辦法傳給其他人） Switch 2 看到了這個封包，想說是廣播封包 所以複製了一份，傳給 Switch 1 與 Switch 3，這時候整個網路拓樸中就已經有兩份這個封包了，讓我們根據複製出來的封包做編號並分別講解

Aug 22, 2021
Read more from HPC-AI-group
Published on HackMD