Hi,
On Mon, 8 Aug 2005, David A. Case wrote:
> On Mon, Aug 08, 2005, John Mongan wrote:
> > Has anyone else tried the new version 9 intel compilers? In my fairly
> > informal testing, sander compiled with ifort 9 seems to pass tests and
> > appears to run 1-2% faster on my (somewhat aging) 2GHz Xeon.
>
> Assuming that the comparison is to ifc 8.x, that would go along with our
> experience: I haven't seen bugs in ifc9/icc9 that would affect sander
or NAB,
> but no great performance increase either.
I have had problems with ifort 8.1 on COLUMBUS and NWChem, such as:
1. the ifort 8.1 fails with a seg fault:
Intel(R) Fortran Compiler for 32-bit applications, Version 8.1 Build
20040921Z Package ID: l_fc_pc_8.1.019
Intel(R) C++ Compiler for 32-bit applications, Version 8.1 Build
20040921Z
Package ID: l_cc_pc_8.1.022
ifort -c -O -DLINUX -DVERIFY -DSTATS maf.F
make[2]: *** [maf.o] Segmentation fault
If one changes the order of the compiler options then the problem
disappears. I have been fearful of trying the 9 version.
> > While on the topic of performance, I've been meaning to ask: has
anyone
> > compared performance between different MPI implementations on
"commodity
> > clusters" (that is, fast dual processor x86 machines with GigE
> > interconnects)? Most people around here seem to use mpich, but I've
been
> > wondering if mpich2 or lam (or something else?) might be better.
>
> lam has been more stable than mpich, so we tend to favor it, not seeing
much
> performance difference. I would hope that people would try mpich2
seriously
> (I've just played with it). It should certainly replace mpich itself,
and
> maybe we could standardize on this, perhaps(?) as a first step towards
> starting to use MPI-2 functionality.
Way back when in sunny san diego, the MP_Lite partial implementation of
MPI
was sometimes faster than either LAM or MPICH, but not by a lot.
I havent tried to build anything with MP_lite recently and I think some
of the later additions to Amber 8 may have used features not in MP_Lite.
I have a dream (or a nightmare) of building Global Arrays on top of
MP_Lite
because of MP_Lite's useful tracing features.
Scott
Received on Wed Apr 05 2006 - 23:49:53 PDT
