On Wed, 25 Feb 2004, Jeff Squyres wrote:
> There is a solution for exactly this problem at the CVS HEAD (other
> PBS users asked for this a few months ago :-) -- a new [as yet
> undocumented] feature called a "hostmap".
OK, that's why I didn't find anything into 7.0.x :-) I'm obviously
looking into 7.0.x as I have lots of binaries compiled/linked with
it...
> Host mapping functionality can be applied to all non-out-of-band
> communication (i.e., MPI communication -- not native LAM/nsend-based
> communication). This nicely fits the "slow/admin" and
> "fast/parallel" network model.
That's exactly what I'm looking for. This is especially more important
as a queueing system would nomally also be configured on the "slow"
network and so lamd could only be started on the "slow" network. So, I
want this feature NOW :-)
As you've put it into 7.1-cvs, maybe you can tell me how difficult
would be to backport it to 7.0. Would a modified lamboot be sufficient
or it has to be present in the parallel program binaries as well ?
> There is now a new SSI parameter "mpi_hostmap" (its prefix of "mpi"
> is meant to imply that it applies to all MPI SSI modules).
I was thinking even farther than this, to allow a command that
receives hostnames on stdin and writes mangled ones on stdout. The
current, file-based behaviour, could then be emulated through 'cat
hostmap'. Then, instead of writting a hostmap, I would use something
like 'sed s/node/gige/'. Of course, this would add more complexity
into LAM...
--
Bogdan Costescu
IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu_at_[hidden]
|
