[kwlug-disc] benchmark: performance on six cores = 1 on one core = 6

Lori Paniak ldpaniak at fourpisolutions.com
Wed Jun 2 22:17:30 EDT 2010


On Wed, 2010-06-02 at 20:52 -0400, John Van Ostrand wrote:
> ----- Original Message -----
> > Not that surprising. Modern CPU cores are the fastest component in a
> > system - by an order of magnitude. Once you get outside of the L2
> > cache, performance takes a big hit. If a job will fit in one core
> > +on-die memory, it will run much faster than the same job spread over
> > several fractionally used cores.
> > 
> > Too bad the whole internet does not yet fit in a single CPU core. If
> > it did, then his test program would model actual usage.
> > 
> > And for high-performance couputing, turn off the hyperthreading. It
> > just confuses things.
> 
> I would have thought that affinity algorithms would have made a process prefer the same core and so context switching would be insignificant.
> 

If the computation is really pushing the limits of the CPU core,
hyperthreading is not going to improve performance.  Plus, you have the
overhead of calculating what to do at each time-slice.  In principle, HT
is a good idea and it works well for most applications, but when each
real core is at 99%+, it just gets in the way.  Put it another way: if
HT helps, you are not doing a good job of keeping the cores busy.

> Another factor is that some new CPUs can overclock one core if the other cores are unused. I think this might have been a BIOS setting rather than a run-time decision.
> 

You can get a pretty good single-core pop from that kind of scaling: 40%
or more.

This is another variable that can mess up a message-passing environment.
If you are running a computation across several machines over a fast
network, timing on the order of a microsecond is critical.  The state of
the system at peak performance is a resonance where all the components
are constructively "aligned".  This is much more difficult to achieve
when each BIOS decides to arbitrarily adjust the core clocks in each
box.

Dynamic clocking:  good for laptops, bad for HPC. 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: This is a digitally signed message part
URL: <http://astoria.ccjclearline.com/pipermail/kwlug-disc_kwlug.org/attachments/20100602/9f66a161/attachment.bin>


More information about the kwlug-disc_kwlug.org mailing list