I'd really like to use mpiexec on amd64, but it looks like the OpenPBS install barfs due to architecture. ---Snip--- checking host system type... Invalid configuration `x86_64-pc-linux-gnuoldld': machine `x86_64-pc' not recognized Invalid configuration `x86_64-pc-linux': machine `x86_64-pc' not recognized configure: error: config.sub failed on x86_64-pc-linux ---Snip--- Torque, however, compiles and runs quite nicely, despite the ~amd64 flag. I'm currently looking into simply writing my own ebuild (and upload it, of course) that will make mpiexec work with torque on AMD64, but I'd love to know if this (torque + mpiexec, x86 or AMD64) has been tried and get advice from those who have tried it. If someone could point me to any notes from previous attempts, that would be great! Also, It seems like SuSE has OpenPBS for x86_64 (http://rpmfind.net/linux/RPM/suse/9.2/i386/suse/x86_64/OpenPBS-clients-2.3.16-604.1.x86_64.html), AND they distribute it (OpenPBS has a Fetch restriction on Gentoo). Any ideas how they do that both in terms of licensing and architecture?
I think we could probably redistribute it as well. Here's the license clauses: 2. Redistribution in any form is only permitted for non-commercial, non-profit purposes. There can be no charge for the Software or any software incorporating the Software. Further, there can be no expectation of revenue generated as a consequence of redistributing the Software. 3. Any Redistribution of source code must retain the above copyright notice and the acknowledgment contained in paragraph 6, this list of conditions and the disclaimer contained in paragraph 7. Probably the reason we don't is that it can't be automatically fetched from the openpbs site.
Okay, so this actually already works. I'm running 1 master node and 8 2-cpu compute nodes (All xen domUs on a 2cpu Opteron physical box) on using Gentoo 2005.1 for AMD64. I've tested with NASA's NPT benchmarks and everything seems to work fine except for the -I flag to qsub (which I think is a torque bug). Could someone please add ~amd64 to sys-cluster/mpiexec-0.76-r2. Maybe if we get some more users we'll discover (and fix) some more problems. :)
Created attachment 79323 [details] virtual libc required by mpiexec ebuild
Created attachment 79324 [details] virtual mpi required by mpiexec ebuild
Created attachment 79325 [details] virtual pbs required by mpiexec ebuild
Ha ha, just kidding. Please delete all the virtual packages I just uploaded as they're not needed. I swear it didn't work without them before, now now it works fine. For those unfamiliar with virtual packages (as I was), they're taken care of by a PROVIDE line in some other package, such as sys-cluster/mpich, sys-libs/glibc, sys-cluster/torque.
See comment #2.
(In reply to comment #7) > See comment #2. > Huh? Nevermind about the binaries. What I'm saying is that I emerged torque (which provides virtual/pbs) and added '=sys-cluster/mpiexec-0.76-r2 ~x86' to packages.keywords on my AMD64 box, and everything seems to work fine, exactly like it should for x86 and ppc, compiling from sources and all. I just wanted you to add ~amd64 to KEYWORDS and possibly some DEPEND logic to require sys-cluster/torque instead of virtual/pbs for AMD64 so we get some more people trying it out.
fails multilib-strict, feel free to reopen with a patch
