[Wien] Hardware Benchmarks for lapw1c

[Alexander Shaposhnikov]

No, export USE_REMOTE=0 does not work either.
I think i have to recompile without mpi support.

This is very bad, as using BLAS-level parallelism i can 
only get no more than 100% speedup. 

I have tested PWScf (pw.x and cp.x), cpmd, and NWChem  codes on this machine.
The Car-Parinello simulation jobs (cpmd, cp.x) seems to have the lowest 
parallel  speedup, as low as 2.5-3 times. Pw.x and nwchem jobs have 3X-5X 
parallel speedup. All jobs with 1 k-pont  only ("fine-grained parallelism").

I have also compared  this machine (2 Quad-core cpus, 8cores) with quad 
Itanium2 1.5GHz (4 CPUS). 
Xeon workstation is 70%-100% faster in MPI jobs  (2.5-3X faster in synthetic 
tests). The latter is also several times less expensive. Itanium seems to 
have very high efficiency, but this cant help to overcome low peak 
floating-point performance (24Gigaflops only for the above-mentioned Quad 
Itanium2).

Best Regards,
Alexander Shaposhnikov

On Friday 26 January 2007 16:01, Florent Boucher wrote:
> Dear Alexander,
> I was not thinking in using the MPI version but just the parallel
> script. You have to change of course the .machines files in order to put
> in the number of different process you want to start.
> May be you should correctly define the USE_REMOTE option in the
> parallel_options file located in $WIENROOT
> You should have
> setenv USE_REMOTE 0
> Regards
> Florent