jnt7 wrote:
>Does anyone have experience with creating 32-bit versions of LAM-MPI on these
>two types of clusters?
>
>AMD Opteron with 64-bit OS Red Hat Enterprise Linux 4 (Portland Group
>compilers v6.0)
>
>
As you would expect, the procedure on the x86-64 OS is much the same
regardless of which brand of compiler or platform is involved. The
32-bit version is done the same as 64-bit, except that all compilers and
linkers are 32-bit. When last discussed in detail here, there was a
hint that all this would soon appear in FAQ, so I won't belabor it. On
EL4, the oldest version of lam which has worked for me is 7.0.3, and the
--enable-shared option has worked well. We gave up on EL4 on clusters
for the time being, due to insufficient Infiniband support. It worked
well with e1000, particularly with interrupt moderation disabled, and
the memory buffer sizes adjusted. The main incentive for using EL4 on
x86-64 is to support multiple dual core CPUs per node, in which case
SuSE is the only alternative.
I don't think any MPI for x86-64 has a way to sense automatically
whether to run in 32- or 64-bit mode, which is among the features
promised for proprietary mpich2 on Windows. If you mix them up, your
processes typically hang suspended, with no diagnostics.
The RHEL4 binutils bugs haven't bitten me on x86-64. I don't know that
anyone has made a list of which EL4 supported platforms will require
their binutils to be replaced by the current public one from kernel.org,
or if there has been any move from RH to correct it. Likewise, I
haven't heard any commitment to fix the broken timers on EL4 x86-64.
|
