By using non-blocking communications, you choose to expose separate
initiation and synchronization MPI calls such that an MPI
implementation is free to schedule the communication in any way it
wants between these two points (while retaining MPI semantics). At
your advantage, this may mean that the MPI implementation can
co-opt specialized hardware/firmware in offloading parts (or all)
of the communication while you get back to computation (or
communication to another processor). If there's nothing to overlap
in your application or if the MPI application has no way to offload
any parts of the communication, all you see is the added cost of
turning one call into two calls. However, with any decent MPI
implementation, this added cost should be a matter of microseconds
(if not nanoseconds), so nothing to worry about for any
non-microbenchmark and at any reasonable scale.
You should also know that since MPI requires local completion only
in that the send buffer be reusable once an MPI send is "complete",
it's perfectly valid for an implementation to simply copy the send
data and claim completion while it defers but guarantees delivery
of the message at a later time. When comparing blocking against
non-blocking, there's probably nothing wrong with the
implementation unless there's a dramatic difference. MPI
implementations play different games on different types of messages
and at different sizes and thresholds. Microbenchmarks just muddy
these issues up even further. All uninteresting stuff, really.
For what you call the MPI broker, it turns out that it's fairly
difficult to completely and cheaply offload an MPI message since
MPI messages have data transfer and implied point-to-point
synchronization (this tags, ranks and communicators business). A
lot of hardware will give you fast primitives for pure data
transfer but none can completely eliminate the cost of brokering or
synchronizing the message from the main application. As a separate
software thread, the broker is tied to scheduling policies and may
positively or negatively compete with the main application thread.
When burried as a helper thread in firmware, the broker has to
negotiate synchronizations at a fraction of the speed of the main
host processor and may not provide sufficient concurrency for your
many-core machine.
. . christian
On Mon, 15 Oct 2007, Eric Thibodeau wrote:
> George,
>
> For completedness's sake, from what I understand here, the
> only way to get "true" communications and computation overlap
> is to have and "MPI broker" thread which would take care of
> all communications in the form of sync MPI calls. It is that
> thread which you call asynchronously and then let it manage
> the communications in the back... correct?
>
> Eric
>
> Le October 15, 2007, George Bosilca a écrit :
> > Eric,
> >
> > No there is no documentation about this on Open MPI. However, what I
> > described here, is not related to Open MPI, it's a general problem
> > with most/all MPI libraries. There are multiple scenarios where non
> > blocking communications can improve the overall performance of a
> > parallel application. But, in general, the reason is related to
> > overlapping communications with computations, or communications with
> > communications.
> >
> > The problem is that using non blocking will increase the critical
> > path compared with blocking, which usually never help at improving
> > performance. Now I'll explain the real reason behind that. The REAL
> > problem is that usually a MPI library cannot make progress while the
> > application is not in an MPI call. Therefore, as soon as the MPI
> > library return after posting the non-blocking send, no progress is
> > possible on that send until the user goes back in the MPI library. If
> > you compare this with the case of a blocking send, there the library
> > do not return until the data is pushed on the network buffers, i.e.
> > the library is the one in control until the send is completed.
> >
> > Thanks,
> > george.
> >
> > On Oct 15, 2007, at 2:23 PM, Eric Thibodeau wrote:
> >
> > > Hello George,
> > >
> > > What you're saying here is very interesting. I am presently
> > > profiling communication patterns for Parallel Genetic Algorithms
> > > and could not figure out why the async versions tended to be worst
> > > than the sync counterpart (imho, that was counter-intuitive). What
> > > you're basically saying here is that the async communications
> > > actually add some sychronization overhead that can only be
> > > compensated if the application overlaps computation with the async
> > > communications? Is there some "official" reference/documentation to
> > > this behaviour from OpenMPI (I know the MPI standard doesn't define
> > > the actual implementation of the communications and therefore lets
> > > the implementer do as he pleases).
> > >
> > > Thanks,
> > >
> > > Eric
> > >
> > > Le October 15, 2007, George Bosilca a écrit :
> > >> Your conclusion is not necessarily/always true. The MPI_Isend is just
> > >> the non blocking version of the send operation. As one can imagine, a
> > >> MPI_Isend + MPI_Wait increase the execution path [inside the MPI
> > >> library] compared with any blocking point-to-point communication,
> > >> leading to worst performances. The main interest of the MPI_Isend
> > >> operation is the possible overlap of computation with communications,
> > >> or the possible overlap between multiple communications.
> > >>
> > >> However, depending on the size of the message this might not be true.
> > >> For large messages, in order to keep the memory usage on the receiver
> > >> at a reasonable level, a rendezvous protocol is used. The sender
> > >> [after sending a small packet] wait until the receiver confirm the
> > >> message exchange (i.e. the corresponding receive operation has been
> > >> posted) to send the large data. Using MPI_Isend can lead to longer
> > >> execution times, as the real transfer will be delayed until the
> > >> program enter in the next MPI call.
> > >>
> > >> In general, using non-blocking operations can improve the performance
> > >> of the application, if and only if the application is carefully
> > >> crafted.
> > >>
> > >> george.
> > >>
> > >> On Oct 14, 2007, at 2:38 PM, Jeremias Spiegel wrote:
> > >>
> > >>> Hi,
> > >>> I'm working with Open-Mpi on an infiniband-cluster and have some
> > >>> strange
> > >>> effect when using MPI_Isend(). To my understanding this should
> > >>> always be
> > >>> quicker than MPI_Send() and MPI_Ssend(), yet in my program both
> > >>> MPI_Send()
> > >>> and MPI_Ssend() reproducably perform quicker than SSend(). Is there
> > >>> something
> > >>> obvious I'm missing?
> > >>>
> > >>> Regards,
> > >>> Jeremias
> > >>> _______________________________________________
> > >>> users mailing list
> > >>> us...@open-mpi.org
> > >>> http://www.open-mpi.org/mailman/listinfo.cgi/users
> > >>
> > >>
> > >
> > >
> > >
> > > --
> > > Eric Thibodeau
> > > Neural Bucket Solutions Inc.
> > > T. (514) 736-1436
> > > C. (514) 710-0517
> >
> >
>
>
>
> --
> Eric Thibodeau
> Neural Bucket Solutions Inc.
> T. (514) 736-1436
> C. (514) 710-0517
>
> _______________________________________________
> users mailing list
> us...@open-mpi.org
> http://www.open-mpi.org/mailman/listinfo.cgi/users
--
christian.b...@qlogic.com
(QLogic Host Solutions Group, formerly Pathscale)
