[mvapich-discuss] mvapich2/1.6 on new AMD Abu Dhabi nodes

[Wesley Emeneker]

Dear MVAPICH developers,
  We have an interesting situation involving old MVAPICH2 versions and very
new hardware.
The basic problem is that the hwloc library (version 1.2) crashes when run
on our new AMD Abu Dhabi nodes.
This affects our installation of mvapich2-1.6 (which has been the default
MPI for the last couple of years) which was compiled against hwloc-1.2.
Updating hwloc to a newer version (1.5 or better seems to work), but when I
try to compile mvapich2-1.6 against the newer version I get this error:
../bin/mpicc   -o cpi cpi.o  -lm
/root/mvapich2-1.6/lib/libmpich.a(ch3_hwloc_bind.o): In function
`smpi_setaffinity':
ch3_hwloc_bind.c:(.text+0x2b99): undefined reference to `hwloc_cpuset_alloc'
ch3_hwloc_bind.c:(.text+0x2f2f): undefined reference to `hwloc_cpuset_set'
ch3_hwloc_bind.c:(.text+0x2fac): undefined reference to `hwloc_cpuset_set'
ch3_hwloc_bind.c:(.text+0x2fdb): undefined reference to `hwloc_cpuset_cpu'
ch3_hwloc_bind.c:(.text+0x302f): undefined reference to `hwloc_cpuset_cpu'


mvapich2-1.9 does not have this problem (crashing in the hwloc library
call).

We would like it if we didn't have to deprecate mvapich2-1.6 for all of our
applications and users.
My question is: can we patch mvapich2-1.6 so that we don't have to go
through the forced-upgrade for everyone?

The hwloc changelog suggests that we can simply replace hwloc_cpusest_ with
hwloc_bitmap_ functions.
Is that correct?

I hope I have clearly described our problem and desired solution.
Let me know if I need to clarify anything.

Thanks,
Wesley


-- 
Wesley Emeneker, Research Scientist
The Partnership for an Advanced Computing Environment
Georgia Institute of Technology

404.385.2303
Wesley.Emeneker at oit.gatech.edu
http://pace.gatech.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20130624/45121de3/attachment.html