[Pw_forum] P4HT problem?

[Serguei Patchkovskii]

On Fri, 9 Jul 2004, Paolo Giannozzi wrote:

> On Friday 09 July 2004 09:11, Dariusz Chrobak wrote:
>
> > Do you have experience with running paraller version of the pw.x
> > on P4 HT processor? [...] CPU time is nearly equal to the CPU time
> > for sequential version of pw.x. Do you have any comments?
>
> Lemon alert:
>
> some multiprocessor motherboards for Intel CPUs have just one
> memory bus for all processors. This dramatically slows down any
> code doing  massive access to memory (such as PWscf and other
> plane-wave codes) that runs on processors belonging on the same
> motherboard.

While this is quite true (in fact, -all- dual-P4 motherboards share
the same front-side bus, which also runs at a slower speed than a
single-P4 FSB - typically 533MHz effective vs. 800), there is likely
another reason for the results seen by Dariusz.

A P4 with "hyperthreading" enabled is simply one CPU, pretending
to be two CPUs, not a proper dual-CPU system. Enabling hyperthreading
on a P4-class CPU does not add any additional computational resources,
it just splits the existing resources differently. It does bring some
benefits to multi-user systems, by allowing two threads of execution
with -dissimilar- resource requirements to share these resources,
instead of taking turns at using them.

Running two execution threads with -similar- resource requirements
on a P4 HT CPU will simply make each one of them run at half the
speed. For a typical high-performance computing workload, the benefits
of HT are usually marginal, or even negative.

Serguei