On Mon, 22 Nov 2004 18:00:20 +0800, Winelfred G. Pasamba
<[EMAIL PROTECTED]> wrote:
> we benchmarked a (P4 3.2 HT, 4GB ram, postgresql, redhat 9) with the
> following configs
> HT disabled, smp kernel;
> HT enabled, smp kernel;
> HT disabled, non-smp kernel;
> HT enabled, non-smp kernel;
> 
> and the fastest is the HT disabled non-smp kernel.
> we saw two cpus when HT was on in the smp kernel. postgresql was using
> both cpus. but performance was a slightly less.
> 
> my *guess* is that HT is fake-smp.  and the slowdown is caused by
> 1) the SMPing (routines, syncs, etc) overhead in the OS level and
> 2) the P4 overhead caused by pretending to be 2 cpus

It is indeed fake. Intel's "HT" is just on-chip simultaneous
multi-threading (SMT), wherein a core can accept more than one kernel
thread.

The slowdown can be attributed to the scheduling algorithms (given
it's not real SMP), as well as hacks that have to be employed given a
flaw in the P4 (and Xeon's) design.

Afaik, the UltraSPARC IV's also have SMT, with the new ones accepting
four kernel threads per core.

-- 
Paolo Alexis Falcone
[EMAIL PROTECTED]
--
Philippine Linux Users' Group (PLUG) Mailing List
[EMAIL PROTECTED] (#PLUG @ irc.free.net.ph)
Official Website: http://plug.linux.org.ph
Searchable Archives: http://marc.free.net.ph
.
To leave, go to http://lists.q-linux.com/mailman/listinfo/plug
.
Are you a Linux newbie? To join the newbie list, go to
http://lists.q-linux.com/mailman/listinfo/ph-linux-newbie

Reply via email to