From: Adrian Knoth (adi_at_[hidden])
Date: 2007-05-19 03:40:41
On Fri, May 18, 2007 at 05:53:21PM -0400, George Bosilca wrote:
[bind BTL/TCP+OOB to specific port]
> As the linux kernel need some time before completely cleaning up the
> socket, this approach can lead to many problems.
Absolutely. My propose cannot be consired useful for productive
environments. As you've already mentioned, it wouldn't be possible to
run more than one process per node.
I see two additional approaches: let the BTL component write its
listening output to stdout (or a file) and collect these information
from every participating node. This is more or less a very reduced
version of WANT_PEER_DUMP.
The second option would dump the corresponding GPR entries, e.g. on
rank 0. I'm thinking of btl_tcp_proc.c:111, somewhere after
mca_pml_base_modex_recv(). In line 144, we iterate over each received
address. It shouldn't be too hard for "Code Master" to write the
ports into a file.
-- Cluster and Metacomputing Working Group Friedrich-Schiller-Universität Jena, Germany private: http://adi.thur.de