Thank you for your reminder, Jeff. This is exactly what I am doing now. I am
avoiding overwriting the buffer for a sending request that has not finished.
I did not observe any message losses so far after I fixed a couple of bugs in
my code.
- Guanhua
On Thursday 05 May 2005 12:04, Jeff Squyres wrote:
> Keep in mind that you can have lots of messages "on the wire" at any
> given time -- they just can't be coming from the same buffer when MPI
> thinks that the message has not completed sending yet (which makes
> total sense -- if you start overwriting your buffer before the message
> has been fully sent [i.e., before MPI_Test* says that it has
> completed], you can't guarantee what is actually sent out).
>
> So you don't have to limit yourself to one MPI_Isend/MPI_Test at a time
> -- but you may have to have a pool of buffers to send from (perhaps
> returning them to the "free" pool when MPI_Test* says that a given
> request has completed).
>
> On May 5, 2005, at 8:32 AM, Guanhua Yan wrote:
> > Thank you a lot for your help, Dave. In my old code, I used MPI_Isend
> > to push
> > out messages aggressively but observed that some messages get lost
> > strangely.
> > So I used a flow control mechanism in which at any time, there is only
> > one
> > message on the fly. MPI_Test is used to test whether the current
> > MPI_Isend is
> > finished. If it's done, the next message will be sent out. However, it
> > seems
> > this approach slows down the throughput significantly. I will roll
> > back to
> > the old version to see whether some bugs caused message losses.
> >
> > - Guanhua
> >
> > On Thursday 05 May 2005 06:38, David Cronk wrote:
> >> There are 2 issues you need to be aware of. First, unless you know a
> >> message has been copied out of the send buffer you cannot touch the
> >> send
> >> buffer. Touching the buffer before the data has been copied out (for
> >> either reading or writing) is strictly prohibited by the standard.
> >> There are only 2 ways to know the message has been copied out of the
> >> send buffer. One is by completing the Isend (with a wait or test
> >> operation). The other is for the receiver to report that the message
> >> has been received.
> >>
> >> The other issue is the request handle returned from the Isend routine.
> >> There are a finit number of these handles available. If you do not
> >> free
> >> these handles (by completing a wait, a test that returns true, or
> >> explicitely with MPI_Request_free) you will eventually run out of
> >> available handles and your app will die on the Isend. You may not run
> >> into this for short runs, but the limit exists and you should free
> >> requests.
> >>
> >> To answer your real question, there should be (as long as the
> >> implementation is standard compliant) no worry about buffer overflow.
> >> What may happen is, if the system runs out of system resources your
> >> send
> >> data will remain in the send buffer until enough system resources have
> >> been freed or the message is matched to a send. See point 1 above for
> >> the conseequences of this.
> >>
> >> I hope this helps.
> >>
> >> Dave.
> >>
> >> Guanhua Yan wrote:
> >>> Hi all,
> >>>
> >>> Sorry if this is off topic. I used MPI_Isend (to the same
> >>> destination)
> >>> for many times without checking whether each sending operation has
> >>> finished. Is it possible that this will run into buffer overflow
> >>> problem
> >>> at the receiver side? Should I do message control by myself or the
> >>> MPI
> >>> kernell handles all these? Or is it MPI implementation dependent?
> >>>
> >>> Thanks,
> >>> Guanhua
> >>> _______________________________________________
> >>> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
> >
> > _______________________________________________
> > This list is archived at http://www.lam-mpi.org/MailArchives/lam/
|
