On Dec 2, 2010, at 3:59 AM, éå£å² wrote:
> When I use openmpi mpirun --mca btl <arg1>, I find arg1 can be ofud, self, sm, openib, but www.open-mpi.org desn't explain those args.
"BTL" stands for "byte transfer layer" -- is the lowest networking software layer for the "ob1" MPI transport in Open MPI (ob1 is usually the default transport in Open MPI).
Each BTL supports a different kind of network:
- ofud: experimental UD-based OpenFabrics transport. I would not use this; it was developed as part of research and was never really finished.
- self: send-to-self (i.e., loopback to the same MPI process)
- sm: shared memory
- openib: generalized OpenFabrics transport.
Open MPI will automatically pick which BTL to use on a per-communication basis, based on which MPI process peer you are communicating with.
The "--mca btl ..." argument to mpirun restricts which BTLs Open MPI will use at run-time.
> I can't understand the mean of "ofud", what different between "ofud" and "openib",
> I also can't understand the different between "ibcm" and "rdmacm", when I use mpirun --mca btl_openib_cpc_include <arg2>.
There are 4 different ways for openib BTL to make connections across Open Fabrics networks:
- oob: the default ("out of band", meaning that it uses TCP sockets)
- xoob: the default when using Mellanox XRC ("out of band with XRC support")
- rdmacm: the default when using iWARP (because iWARP doesn't support OOB or XOOB)
- ibcm: not currently used; it's an IB-specific method that was never really finished
Usually, the right CM is just automatically picked -- you shouldn't need to manually select anything.
> maybe www.open-mpi.org can publish a openmpi's document to explain those args and principle.
We are lacking in the documentation department; contributions would be greatly appreciated...
The README file has a bunch about BTLs; that may be helpful reading.
--
Jeff Squyres
jsquyres_at_[hidden]
For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/
|