Presentation is loading. Please wait.

Presentation is loading. Please wait.

MPI and RDMA Yufei 10/15/2010. MPI over uDAPL: abstract MPI: most popular parallel computing standard MPI needs the ability to deliver high performace.

Similar presentations


Presentation on theme: "MPI and RDMA Yufei 10/15/2010. MPI over uDAPL: abstract MPI: most popular parallel computing standard MPI needs the ability to deliver high performace."— Presentation transcript:

1 MPI and RDMA Yufei 10/15/2010

2

3 MPI over uDAPL: abstract MPI: most popular parallel computing standard MPI needs the ability to deliver high performace Implement MPI with both portability and portable high performance using uDAPL. Two important techs: – Improve the startup time and communication performance – Multi-stream MPI design

4 MPI over uDAPL: background InfiniBand Vitual Interface Architecture iWARP uDAPL MPI MPI applicationapplication

5 MPI over uDAPL: Efficient Connection Establishment Multi-threads vs. Non-thread n processes need establish n*(n-1)/2 connections. If each connection needs a thread … – Thread creation and switching overhead – Portable problem of thread How to synchronize each process and step?

6 MPI over uDAPL: Multi-Stream MPI Design Deal with out-of-order messages: sequence number

7 MPI over uDAPL: Memory Management Should all the EPs share the same set of RDMA buffers or should they have separate buffers? But how could different streams share the buffers?

8 MPI over uDAPL: Connection Selection Policy

9 MPI over uDAPL: Result: threadless connection establishment

10 MPI over uDAPL: Result: bandwidth over infiniband on 10Gb

11

12 RDMA-Based MPI: Introduction InfiniBand Architecture – Channel semantics (communication) – Memory semantics (RDMA READ/WRITE) MPI is the standard for writing parallel apps Key techs – Persistent buffer association – RDMA polling set

13 RDMA-Based MPI: Send/Receive vs. RDMA Write Latency shows RDMA Writer is better than Send/Receive?

14 RDMA-Based MPI: Use RDMA Write instead of Send/Receive RDMA Write designed to implement Control Message – The RDMA destination address must be known before communication Persistent buffer association – The receiver side must detect the arrival of incoming message RDMA polling set

15 RDMA-Based MPI: Persistent buffer association

16 RDMA-Based MPI: RDMA polling set But how to poll the RDMA Polling Set (or Memory Region)

17 RDMA-Based MPI: RDMA Buffer Structure Waste CPU or Waste Time?

18 Q & A


Download ppt "MPI and RDMA Yufei 10/15/2010. MPI over uDAPL: abstract MPI: most popular parallel computing standard MPI needs the ability to deliver high performace."

Similar presentations


Ads by Google