[mpich-discuss] Active loop in MPI_Waitany?
Dorier, Matthieu
mdorier at anl.gov
Thu Jul 30 18:48:58 CDT 2015
Not that obvious, a UNIX function like "select" would allow to block on a set of file descriptors (for instance sockets), so I was expecting MPI_Waitany to work kind of the same way. Pavan explained me (in person) why it does not: it cannot at the same time block on a read from shared memory (channels inside a node) and on network connections (outside the node). (I'm simplifying a bit here).
So the answer is: yes, MPI_Waitany is actively looping.
Matthieu
________________________________
From: Jeff Hammond [jeff.science at gmail.com]
Sent: Thursday, July 30, 2015 6:41 PM
To: discuss at mpich.org
Subject: Re: [mpich-discuss] Active loop in MPI_Waitany?
Seems obvious that Waitany spins on the array of requests until one completes. Is that an active loop by your definition?
Jeff
On Thursday, July 30, 2015, Dorier, Matthieu <mdorier at anl.gov<mailto:mdorier at anl.gov>> wrote:
Hi,
I have a code that looks like this:
while(true) {
do some I/O (HDF5 POSIX output to a remote, parallel file system)
wait for communication (MPI_Waitany) from other processes (in the same node and outside the node)
}
I'm measuring the energy consumption of the node that runs this process for the same duration, as a function of the amount of data written in each I/O operation.
Surprisingly, the larger the I/O in proposition to the communication, the lower the energy consumption. In other words, the longer I wait in MPI_Waitany, the more I consume.
Does anyone have a good explanation for that? Is there an active loop in MPI_Waitany? Another reason?
Thanks!
Matthieu
--
Jeff Hammond
jeff.science at gmail.com<mailto:jeff.science at gmail.com>
http://jeffhammond.github.io/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20150730/378963a3/attachment.html>
-------------- next part --------------
_______________________________________________
discuss mailing list discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss
More information about the discuss
mailing list