Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

From: Tim S. Woodall (twoodall_at_[hidden])
Date: 2005-11-10 11:10:56


Mike,

I believe this issue has been corrected on the trunk, and should
be in the next release candidate, probably by the end of the week.

Thanks,
Tim

Mike Houston wrote:
> mpirun -mca btl_mvapi_rd_min 128 -mca btl_mvapi_rd_max 256 -np 2
> -hostfile /u/mhouston/mpihosts mpi_bandwidth 21 131072
>
> 131072 519.922184 (MillionBytes/sec) 495.836433(MegaBytes/sec)
>
> mpirun -mca btl_mvapi_rd_min 128 -mca btl_mvapi_rd_max 256 -np 2
> -hostfile /u/mhouston/mpihosts mpi_bandwidth 22 131072
>
> 131072 3.360296 (MillionBytes/sec) 3.204628(MegaBytes/sec)
>
> Moving from 21 messages to 22 causes a HUGE drop in performance. The
> app tries to send all of the messages non-blocking at once... Setting
> -mca mpi_leave_pinned 1 causes:
>
> [0,1,1][btl_mvapi_component.c:631:mca_btl_mvapi_component_progress] Got
> error : VAPI_WR_FLUSH_ERR, Vendor code : 0 Frag : 0xb73412fc
>
> repeated until it eventually hangs.
>
> -Mike
>
> Mike Houston wrote:
>
>
>>Woops, spoke to soon. The performance quoted was not actually going
>>between nodes. Actually using the network with the pinned option gives:
>>
>>[0,1,0][btl_mvapi_component.c:631:mca_btl_mvapi_component_progress]
>>[0,1,1][btl_mvapi_component.c:631:mca_btl_mvapi_component_progress] Got
>>error : VAPI_WR_FLUSH_ERR, Vendor code : 0 Frag : 0xb74a1c18Got error :
>>VAPI_WR_FLUSH_ERR, Vendor code : 0 Frag : 0xb73e1720
>>
>>repeated many times.
>>
>>-Mike
>>
>>Mike Houston wrote:
>>
>>
>>
>>
>>>That seems to work with the pinning option enabled. THANKS!
>>>
>>>Now I'll go back to testing my real code. I'm getting 700MB/s for
>>>messages >=128KB. This is a little bit lower than MVAPICH, 10-20%, but
>>>still pretty darn good. My guess is that I can play with the setting
>>>more to tweak up performance. Now if I can get the tcp layer working,
>>>I'm pretty much good to go.
>>>
>>>Any word on an SDP layer? I can probably modify the tcp layer quickly
>>>to do SDP, but I thought I would ask.
>>>
>>>-Mike
>>>
>>>Tim S. Woodall wrote:
>>>
>>>
>>>
>>>
>>>
>>>
>>>>Hello Mike,
>>>>
>>>>Mike Houston wrote:
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>>When only sending a few messages, we get reasonably good IB performance,
>>>>>~500MB/s (MVAPICH is 850MB/s). However, if I crank the number of
>>>>>messages up, we drop to 3MB/s(!!!). This is with the OSU NBCL
>>>>>mpi_bandwidth test. We are running Mellanox IB Gold 1.8 with 3.3.3
>>>>>firmware on PCI-X (Couger) boards. Everything works with MVAPICH, but
>>>>>we really need the thread support in OpenMPI.
>>>>>
>>>>>Ideas? I noticed there are a plethora of runtime options configurable
>>>>>for mvapi. Do I need to tweak these to get performacne up?
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>>You might try running w/ the:
>>>>
>>>>mpirun -mca mpi_leave_pinned 1
>>>>
>>>>Which will cause mvapi port to maintain an mru cache of registrations,
>>>>rather than dynamically pinning/unpinning memory.
>>>>
>>>>If this does not resolve the BW problems, try increasing the
>>>>resources allocated to each connection:
>>>>
>>>>-mca btl_mvapi_rd_min 128
>>>>-mca btl_mvapi_rd_max 256
>>>>
>>>>Also can you forward me a copy of the test code or a reference to it?
>>>>
>>>>Thanks,
>>>>Tim
>>>>_______________________________________________
>>>>users mailing list
>>>>users_at_[hidden]
>>>>http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>
>>>_______________________________________________
>>>users mailing list
>>>users_at_[hidden]
>>>http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>
>>>
>>>
>>>
>>
>>_______________________________________________
>>users mailing list
>>users_at_[hidden]
>>http://www.open-mpi.org/mailman/listinfo.cgi/users
>>
>>
>
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users
>