- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I would like to run a heterogeneous program on a multi node cluster where each node comes with 2 MIC devices.
Let's consider for example 2 dual-sockets nodes:
------------------- -------------------
--- NODE 1 -- --- NODE 2 --
| |
S | S S | S
o | o o | o
c | c c | c
k | k k | k
e | e e | e
t | t t | t
#1 | #2 #1 | #2
------------------- -------------------
MIC0 | MIC1 MIC0 | MIC1
------------------- -------------------
If I execute 4 mpi processes, how can I assure that each mpi process will run on a different socket, and that each mpi-rank will offload on the MIC device connected to the socket where it is executed?
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi Ale,
Did you try using "mpirun -ppn 2 -np 4 I_MPI_PIN_DOMAIN=socket ....".
Also I am assuming that you have Intel Xeon Phi coprocessors configured in external bridge topology and are able to ping each other across cluster.
Thanks
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page