If you use one of the latest version of MX (such as 1.2.0f) there
will be less than 0.15 micro-second difference between the MTL and
BTL MX. This version of MX allow us to do the matching outside the
NIC, which decrease the overhead for small messages. In terms of
bandwidth the BTL MX is so
Note that since you are setting OMPI_MCA_pml to cm, OMPI_MCA_btl will have no
effect. You may try setting OMPI_MCA_pml=ob1, and trying your measurements
again, but we generally get better performance with the cm pml than then ob1
pml.
Tim
On Wednesday 06 June 2007 12:54:26 pm George Bosilca wr
George,
Apologies for not saying what latency is with Open MPI. I've noted it below.
I don't know why turning off the sm feature would help launching 1ppn. I just
tried turning it off and it didn't make a difference.
-cdm
On Wed, 6 Jun 2007, George Bosilca wrote:
Which one is the latency
Which one is the latency with Open MPI ? Which version of Open MPI ?
You might want to use OMPI_MCA_btl=mx,self to see if it makes any
difference.
Thanks,
george.
On Jun 6, 2007, at 12:26 PM, Maestas, Christopher Daniel wrote:
With 2 nodes using 1.1.7 with the patch we measured (usin
With 2 nodes using 1.1.7 with the patch we measured (using mpich-mx
1.2.7..4):
3.07us
With mx 1.2.1-rc18 we measure:
3.69 us
And with mpich-mx 1.2.7..4 we see:
3.20us
Our Open MPI settings:
---
# env | grep OMPI
OMPI_MCA_pml=cm
OMPI_MCA_mpi_keep_hostnames=1
OMPI_MCA_oob_tc
