[Wien] Hardware Benchmarks for lapw1c

Alexander Shaposhnikov shaposh at isp.nsc.ru
Fri Jan 26 10:52:21 CET 2007


On Friday 26 January 2007 13:58, Florent Boucher wrote:
> Dear Alexander,
> could you please also test a parallel job (just put 2, 4 and 8 lines in
> the klist file with the same kpoint and submit with the parallel option)
> in order to evaluate the how good is the memory band width.
> A perfect system (without saturation of the memory band width) should
> give the same CPU time to calculate 1 k-point, whatever the number of
> parallel job submitted.
> I am expecting rather bad performance for 4 and 8 parallel job on Xeon

I cant figure out how to run it in parallel.
I've compiled mpi version, and it works very weird.

With 1 k-point and 2 mpi threads, the test job finishes in 89sec.
However, with more than 2 threads, it works forever.

With 2 k-points, serial execution goes normally, but  parallel executions
somehow don't use the second k-point.
The  "x" script also never  finishes normally
for parallel execution, i have to ctrl-C it.

Is mpi version broken/not reliable currently?  

> Please could you please be more precise about the frequency of the CPU
> (1.86GHz->2.67GHz)
The machine is dual Xeon Clovertown E5320 with default 1.86GHz frequency.
This is my personal workstation, so i managed to overclock it a bit to 
2.67GHz  :). 
The peak floating point performance  is 85.5 Gigaflops with really attainable 
62.5 Gigaflops (linpack-hpl with Goto BLAS)


 Best Regards,
 Alexander Shaposhnikov



More information about the Wien mailing list