On Dec 31, 2007, at 11:42 PM, Paul H. Hargrove wrote:
I tried today to build the OMPI trunk on a system w/ GM libs installed
(I tried both GM-2.0.16 and GM-1.6.4) and found that the GM BTL won't
even compile, due to unbalanced parens. The patch below reintroduces
the parens that were apparently lost in r16633:
Fixed (https://svn.open-mpi.org/trac/ompi/changeset/17029); thanks for
the patch.
The fact that this has gone unfixed for 2 months suggests to me that
nobody is building the GM BTL. So, how would I go about checking ...
a) ...if there exists any periodic build of the GM BTL via MTT?
I thought that Indiana was doing GM builds, but perhaps they've
upgraded to MX these days...?
UTK -- do you still have any GM clusters around?
b) ...if such builds, if any, experience the same error(s) as I
c) ...which GM library versions such builds, if any, compile against
Given the typos you found, I don't see how they could.
d) ...if anybody wants to help setup an MTT for GM on my system (NOTE:
Jeff Squyres, Brian Barrett and George Bosilca all have existing
accounts on my cluster, though possibly expired/disabled).
I always like to see more OMPI testing. :-)
I'd be happy to help setup MTT for your cluster. Is it easy to re-
activate my accounts? What kind of testing would you be willing to do
on your cluster, and how often? What queueing system do you
use? ...etc. (this might be worth a phone call)
I have a somewhat-complex setup for MTT on my Cisco development
cluster; I submit a whole pile of compilation MTT jobs via SLURM and
wait for them to complete (individually). Each compile that completes
successfully will end up generating another pile of SLURM jobs for
testing. I have a [somewhat ugly] top-level script that submits all
these jobs according to a schedule set by day-of-week.
Sidenote: one of the interesting things about MTT that we've found is
that everyone tends to use it differently -- IU, Sun, IBM, and Cisco
all use MTT quite differently in our nightly regression testing. So
our top-level scripting to invoke MTT is not uniform at all. We've
long-since talked about adding a uniform upper layer for large-scale
MTT automation that can handle full parallelism, generic batch queue
system support, etc., but haven't found the cycles to get together and
try to map out what it would look like. Plus, all of our individual
setups are working / ain't broken, so there's not a lot of incentive
to "fix" them... It might be an interesting software engineering
research project, though, if anyone's got the cycles. This has [much]
larger implications than just MPI testing.
--
Jeff Squyres
Cisco Systems