Le 16/12/2010 01:40, Jeff Squyres a écrit :
gg= No, this problem happens rather often, almost every other time.
On Dec 15, 2010, at 3:24 PM, Ralph Castain wrote:
I am not using the TCP BTL, only OPENIB one. Does this change the number of sockets in use per node, please ?
I believe the openib btl opens sockets for connection purposes, so the count is likely the same. An IB person can confirm that...
Nope -- the openib BTL uses the daemon-based communication mechanism. So it should only use the TCP ports that are already open.
Does this problem *always* happen, or does it only happen once in a great while?
Seems to happen more often as the number of cores increases.
gg= Is there a way with the current code, to direct OpenMPI
to use a restricted range of TCP ports,
I've seen a similar problem with the TCP BTL every once in a great while -- where a random, errant (non-Open MPI) process connects to a socket that Open MPI is listening on (regardless of whether it's the TCP BTL or TCP OOB). This causes badness in Open MPI because we don't verify the connector properly, and more importantly, don't handle it nicely when the connector is not Open MPI. I've seen this happen with network malware scanners, for example -- they try to connect to large swaths of TCP ports and sometimes unluckily hit an open Open MPI TCP port.
that I can choose at launch time ?
Or, conversely, which routine should I patch in my private OpenMPI
install to aim at the same result ?
Usually, on the cluster I use, workers are not shared with any
other tasks ... ;;-((
Thanks, Best, G.
We have a fix for this coming in a future version of the TCP BTL; looks like we should also harden this up for the TCP OOB as well...