[Beowulf] Liinpack benchmark

Jakob Oestergaard jakob at unthought.net
Wed Nov 9 01:09:16 PST 2005

On Tue, Nov 01, 2005 at 07:09:39PM +0300, Mikhail Kuzminsky wrote:
> In message from balamurugan <rbala at hclinsys.com> (Mon, 31 Oct 2005 
> 15:06:11 +0530):
> >Dear all,
> >Iam trying to benchmark a Xeon processor (Dual core) machine with 
> >linpack (n=100) fortran code on Red Hat linux AS 4. The result gives 
> >zero values for some calls. The details in the internet say that it 
> >may be due to low resolution of the timing function. Can you help me 
> >tuning the resolution of second function.
> You should use gettimeofday()- based or RDTSC-based timer.

Is the TSC synced between the cores on the Xeon?  If not, then simple
rdtsc based timing is going to be unreliable.  (It's possible of course,
with many repeated tests and outlier elimination to disregard the tests
where the thread switched core).

I know the TSC isn't synced between cores on dual core Opteron, it's not
synced on plain SMP (multiple sigle-core processors) and given that the
dual core Xeon is pretty much just two Xeons shoe-horned into one
package, I doubt it would be synced there.  (But, as always, would love
to be corrected, should I be wrong :)

gettimeofday() is definitely a very good suggestion.


 / jakob

More information about the Beowulf mailing list