Thank you very much for your reply. Unfortunately, the overloading is
not the problem. The phenomenon also appears if we use only two
processes on the 8core machines. When I run the jobs over two nodes, one
is doing nothing anymore after a couple of minutes. The strange thing
is that this only happens on ifiniband and only with mpi2 libraries
(openmpi and mvapich2). Mvapich1 is running reasonably fine at the
moment. Perhaps the first to mpi implementations have something in
common, which could trigger the problems.
Jeff Squyres wrote:
> Sorry for the delay in replying -- INBOX deluge makes me miss emails on
> the users list sometimes.
> I'm unfortunately not familiar with gamess -- have you checked with
> their support lists or documentation?
> Note that Open MPI's IB progression engine will spin hard to make
> progress for message passing. Specifically, if you have processes that
> are "blocking" in message passing calls, those processes will actually
> be spinning trying to make progress (vs. actually blocking in the
> kernel). So if you overload your hosts -- meaning that you run more
> Open MPI jobs than there are cores -- you could well experience dramatic
> slowdown in overall performance because every MPI job will be competing
> for CPU cycles.
> On Feb 24, 2009, at 4:57 AM, Thomas Exner wrote:
>> Dear all:
>> Because I am new to this list, I would like to introduce myself as
>> Thomas Exner and please excuse silly questions, because I am only a
>> And now my problem, with which I am fiddling around for almost a week: I
>> try to use gamess with openmpi on infiniband. There is a good
>> description on how to compile it with mpi and it can be done, even if
>> it is not easy. But then on run time everything gets weird. The
>> specialty of gamess is that it runs twice as much mpi jobs than used for
>> the computation. The second half is used as data server, requiring data
>> but with very little cpu load. Each one of these data servers is
>> connected to a specific compute job. Therefore, these two corresponding
>> jobs have to be run on the same node. On one node everything is fine
>> (2x4core machines in my case), because all the jobs are guarantied to
>> run on this node. If I try two nodes, at the beginning also everything
>> is fine. 8 compute jobs and 8 data server are running on each machine.
>> But after a short while, the entire set of processes (16) on the first
>> node start to accumulate CPU time, with nothing useful happening. The
>> second node's processes go entirely to sleep. Is it possible that all
>> the compute jobs are for some reason been transfered to the first node?
>> This would explain the load of 16 on the first and 0 on the second node,
>> because 16 compute jobs (100 % cpu load) and 16 data servers (almost 0%
>> load) are running, respectively. Strange thing is also that the same
>> version runs on gigabit and myrinet fine.
>> It would be great if somebody could help me on that. If you need more
>> information, I will be happy to share them with you.
>> Thanks very much.
>> users mailing list