more dbt-2 results hyperthreading on linux-2.6.0-test11

On 12 Dec, Nakajima, Jun wrote:

I'm not familiar with this particular workload, but noticed higher idle
and wait time with HT enabled, compared to HT-disabled case. This kind
of symptom often indicated insufficient I/O bandwidth from my
experience, and faster systems (with more threads) tend to show lower
throughput because you end up measuring disk seek time of more I/O
requests. Can you add more disk controller(s) and disks?

Jun

-----Original Message-----
From: linux-kernel-owner@vger.kernel.org [mailto:linux-kernel-
owner@vger.kernel.org] On Behalf Of markw@osdl.org
Sent: Friday, December 12, 2003 2:28 PM
To: piggin@cyberone.com.au
Cc: mingo@redhat.com; linux-kernel@vger.kernel.org; pgsql-
hackers@postgresql.org
Subject: more dbt-2 results hyperthreading on linux-2.6.0-test11

Hi Nick,

Here are the results of the comparisons I said I would do.

no-hyperthreading:
http://developer.osdl.org/markw/dbt2-pgsql/282/
- metric 2288.43
- baseline

hyperthreading:
http://developer.osdl.org/markw/dbt2-pgsql/278/
- metric 1944.42
- 15% throughput decrease

hyperthreading w/ Ingo's C1 patch:
http://developer.osdl.org/markw/dbt2-pgsql/277/
- metric 1978.39
- 13.5% throughput decrease

hyperthreading w/ Nick's w26 patch:
http://developer.osdl.org/markw/dbt2-pgsql/274/
- metric 1955.91
- 14.5% throughput decrease

It looks like there is some marginal benefit to your or Ingo's patches
with a workload like DBT-2. I probably don't understand enough about
hyperthreading, but I wonder if there's something PostgreSQL can do to
take advantage of hyperthreading

Anyway, each link has pointers to readprofile and annotated oprofile
assembly output (if you find that useful.) I haven't done enough

tests

to have an idea of the error margin, but I wouldn't be surprised if

it's

at least 1%.

Let me know if there's anything else you'd like me to try.

Thanks,
Mark