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: 2007-10-23 10:23:11


Currently there is no work around this issue. We consider(ed) that
when you run an MPI job the cluster is in dedicated mode, so a 100%
CPU consumption is acceptable. However, as we discussed at our last
meeting, there are others reasons to be able to yield the CPU until a
message arrives. Therefore, we plan to have a blocking mode in the
near future. The is no timeframe for this, but the discussions
already started (that is usually a good sign).

   Thanks,
     george.

On Oct 23, 2007, at 9:17 AM, Murat Knecht wrote:

> Hi,
> thanks for answering. Unfortunately, I did try that, too. The point
> is that i don't understand the ressource consumption. Even if the
> processor is yielded, it still is busy waiting, wasting system
> resources which could otherwise be used for actual work. Isn't
> there some way to activate an interrupt mechanism, so that the wait/
> recv blocks the thread, e.g. puts it to sleep, until notified?
>
> Murat
>
> Tim Mattox schrieb:
>> You should look at these two FAQ entries: http://www.open-mpi.org/
>> faq/?category=running#oversubscribing http://www.open-mpi.org/faq/?
>> category=running#force-aggressive-degraded To get what you want,
>> you need to force Open MPI to yield the processor rather than be
>> aggressively waiting for a message. On 10/23/07, Murat Knecht
>> <murat.knecht_at_[hidden]> wrote:
>>> Hi, Testing a distributed system locally, I couldn't help but
>>> notice that a blocking MPI_Recv causes 100% CPU load. I
>>> deactivated (at both compile- and run-time) the shared memory bt-
>>> layer, and specified "tcp, self" to be used. Still one core busy.
>>> Even on a distributed system I intend to perform work, while
>>> waiting for incoming requests. For this purpose having one core
>>> busy waiting for requests is uncomfortable to say the least. Does
>>> OpenMPI not use some blocking system call to a tcp port
>>> internally? Since i deactivated the understandably costly shared-
>>> memory waits, this seems weird to me. Someone has an explanation
>>> or even better a fix / workaround / solution ? thanks, Murat
>>> _______________________________________________ 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



  • application/pkcs7-signature attachment: smime.p7s