Open MPI logo

Open MPI User's Mailing List Archives

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

From: George Bosilca (bosilca_at_[hidden])
Date: 2006-08-29 14:07:12


Well, you post the question on the user mailing list. Therefore, I
give you a user answer. If you want to know how do we implemented
deep inside, then the answer is different.

Short messages even on synchronous mode are sent using the rendez-
vous protocol (sic). Except that our rendez-vous protocol transport
some data. How much data, that depend on the network and it can be
modified via an MCA parameter (first_fragment).

In all the cases the for the synchronous send we will wait for an ack
from the remote side before marking the request as completed at the
sender. In fact, internally it happens exactly how you described in
your email.

   george.

On Aug 29, 2006, at 1:32 PM, Marcelo Stival wrote:

> Thanks for your replay...
>
> Let me be a little insistent... :P
>
> I read (I don't remember where...) that an specific implementation
> could send messages with synchronous comm mode... eagerly!
> I think it still conform to the standard if the operation completes
> just after the ack from the receiver.
>
> The MPI_Ssend() could send a short message eagerly and wait for the
> ack... (to satisfy the semantics of synchronous send)
>
> Well, if it's possible.... I want to know if the OPENMPI
> implementation (with BTL TCP over Ethernet) uses different protocols
> for short/long when synchronous ...
> (or it will be always rendezvous... as stated before)
>
>
> Thanks
> Marcelo
>
> On 8/29/06, George Bosilca <bosilca_at_[hidden]> wrote:
>> On Tue, 29 Aug 2006, Marcelo Stival wrote:
>>
>>> Hi,
>>>
>>> I have 2 questions related to short/long message protocols...
>>>
>>> 1) When using synchronous comm mode, short messages (<64kB)
>>> still be
>>> transferred eagerly?
>>
>> Depend. 64Kb is not the limits between the eager and rendez-vous
>> protocol.
>> In fact it depend on the underlying network that get activated.
>> Anyway,
>> for this particular question this limit between the eager and
>> rendez-vous
>> protocol does not matter. Synchronous is alway a rendez-vous
>> protocol as
>> specified in the MPI standard.
>>
>>> And larger messages will be transferred using rendezvous...?
>>
>> That's always the case even for non synchronous operations.
>>
>>>
>>> 2) When the progress-thread (tcp btl) will be applied?
>>> Just for long messages? Does it affect synchronous short messages
>>> sends?
>>
>> The behavior will be the same with or without threads.
>>
>> Thanks,
>> george.
>>
>>
>> "We must accept finite disappointment, but we must never lose
>> infinite
>> hope."
>> Martin Luther King
>>
>> _______________________________________________
>> 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

"Half of what I say is meaningless; but I say it so that the other
half may reach you"
                                   Kahlil Gibran