Look here:

http://www.open-mpi.org/faq/?category=tuning#selecting-components

General idea

mpirun -np 2 --mca btl ^tcp (to exclude ethernet) replace with ^openib (or ^mvapi) to exclude infiniband.

Brock Palen
Center for Advanced Computing
bro...@umich.edu
(734)936-1985


On Apr 18, 2007, at 8:44 AM, stephen mulcahy wrote:

Hi,

I'm currently conducting some testing on system with gigabit and
infiniband interconnects. I'm keen to baseline openmpi over both the
gigabit and infiniband interconnects.

I've compiled it with defaults and run the Intel MPI Benchmarks PingPong as follows to get an idea of latency and bandwidth between nodes on the
given interconnect.

~/openmpi-1.2/bin/mpirun --bynode -np 2 --hostfile ~/openmpi.hosts.80
~/IMB/IMB-MPI1-openmpi -npmin 2 pingpong

For some reason, it looks like openmpi is using the infiniband
interconnect rather than the gigabit ... or the system I'm testing on
has an amazing latency! :)

#---------------------------------------------------
# Benchmarking PingPong
# #processes = 2
#---------------------------------------------------
        #bytes #repetitions      t[usec]   Mbytes/sec
             0         1000         1.63         0.00
             1         1000         1.54         0.62
             2         1000         1.55         1.23
             4         1000         1.54         2.47
             8         1000         1.56         4.90
            16         1000         1.86         8.18
            32         1000         1.94        15.75
            64         1000         1.92        31.77
           128         1000         1.99        61.44
           256         1000         2.25       108.37
           512         1000         2.70       180.88
          1024         1000         3.64       267.99
          2048         1000         5.60       348.89

I read some of the FAQs and noted that OpenMPI prefers the faster
available interconnect. In an effort to force it to use the gigabit
interconnect I ran it as follows,

~/openmpi-1.2/bin/mpirun --mca btl tcp,self --bynode -np 2 --hostfile
~/openmpi.hosts.80 ~/IMB/IMB-MPI1-openmpi -npmin 2 pingpong

and

~/openmpi-1.2/bin/mpirun --mca btl_tcp_if_include eth0 --mca btl
tcp,self --bynode -np 2 --hostfile ~/openmpi.hosts.80
~/IMB/IMB-MPI1-openmpi -npmin 2 pingpong

Neither one resulted in a significantly different benchmark.

Am I doing something obviously wrong in how I invoke openmpi here or
should I expect this to run over gigabit? Is there an option to mpirun
which I can provide to tell me what interconnect it does use?

I gave a look at the ompi_info output but couldn't see any indication
that infiniband support was compiled in so I'm a little puzzled by this
but the results speak for themselves.

Any advice on how to force the use of gigabit would be welcomed (I'll
use the infiniband interconnect aswell but I'm trying to determine the
performance to be had from infiniband for our model so I need to run it
with both).

Thanks,

-stephen
--
Stephen Mulcahy, Applepie Solutions Ltd., Innovation in Business Center, GMIT, Dublin Rd, Galway, Ireland. +353.91.751262 http:// www.aplpi.com
Registered in Ireland (289353) (5 Woodlands Avenue, Renmore, Galway)
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users



Reply via email to