Open MPI logo

MTT Devel Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all MTT Devel mailing list

From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2007-08-29 08:58:36


On Aug 29, 2007, at 7:12 AM, Josh Hursey wrote:

>>>> - We need a well defined way to see what collective implementation
>>>> was used. Meaning that there are N AlltoAll collective
>>>> implementations in the 'tuned' component we need to know when
>>>> looking at the graph which one of the N we are looking at for
>>>> Open MPI. For
>>>> other implementations we don't have so much control.
>>
>> I don't know if MTT can. In order for MTT to do this, OMPI needs to
>> export that data somehow.
>
> So I see 2 solutions to this:
> 1) Require that everyone specify the collective on the command line.
> 2) Create a special, hidden MCA parameter that prints the
> collective being used only once. Then MTT can extract that and we can
> track it.
> Just an idea.

Hmm. But remember that "tuned" may pick a different algorithm for
any given data point on the graph. So for a single skampi/IMB/
whatever Bcast run going from 1 byte to 8MB, you could have lotsa
different algorithms used.

>> http://www.open-mpi.org/mtt/reporter.php?do_redir=290
>>
>> I don't see why this is hard...? I disagree with the statement
>> "Reporter is design to track a single MPI well..." See the permalink
>> above.
>
> Humm I was having an aweful time getting these results to pair as you
> did in the link you gave me (I actually gave up after a while). Maybe
> I was using the reporter wrong.

For me, that URL goes to 2 results: skampi bcast for intel MPI 3.0
and ompi-trunk-r15963 and shows the two respective lines on a latency
graph.

Are you seeing something different? You might need to have trials
enabled? (the two results were trials -- sorry, I didn't mention
that...)

-- 
Jeff Squyres
Cisco Systems