[Wien] wien2k, gotoblas and multi threads

[Peter Blaha]

> I'm confused by the part of the web site benchmarks, here:
> 
>    http://www.wien2k.at/reg_user/benchmark/
> 
> under the section titled "Serial benchmark with parallel jobs".
> What exactly is that section showing?  What factor provides the
> speedup?

I thought this should be "clear" after our previous correspondence.
Anyway:
---------------------
Serial benchmark with parallel jobs (Tests the "real" performance under 
full load with a "k-parallel" job): NMAT=3481, complex

1333 FSB Dual-Clovertown X5355  @ 2.66GHz, 667 Memory					
Jobs   1 Thread    2 Threads   4 Threads   8 Threads
1        132            88          66         62	
2        145           104          98	
4        177           163	
---------------------
If you have 1 k-point (in case.klist) you can run it in 62 seconds 
(using 8 threads).

Suppose you have 4 k-points in case.klist, than running in this mode 
means a time of 4*62=248 sec.

If you run in k-parallel mode, it means the script starts 4 different 
lapw1c jobs simultaneously, thus each "job" does only one k-point.
Unfortunately, current Xeons are not fully balanced systems, but memory 
is too slow. Thus when 4 jobs are executed simultaneously, each of them 
will be SLOWER. Nevertheless, it will take only 163 seconds to finish 
all 4 k-points, thus k-point parallel is 85 sec faster than sequential 
(but multitheading) mode.

Regards

-- 
-----------------------------------------
Peter Blaha
Inst. Materials Chemistry, TU Vienna
Getreidemarkt 9, A-1060 Vienna, Austria
Tel: +43-1-5880115671
Fax: +43-1-5880115698
email: pblaha at theochem.tuwien.ac.at
-----------------------------------------