8rc2 & BLCKSZ

Hi,

A small test with 8rc2 and BLCKSZ of 8k and 32k.
It seems there is a 10% increase in the number of transactions by
second.
Does someone plan to carefully test the impact of BLCKSZ ?

Cordialement,
Jean-Gérard Pailloncy

with 8k:
> /test/bin/pgbench -c 10 -t 300 test
starting vacuum...end.
transaction type: TPC-B (sort of)
scaling factor: 10
number of clients: 10
number of transactions per client: 300
number of transactions actually processed: 3000/3000
...
tps = 26.662146 (excluding connections establishing)
tps = 23.742071 (excluding connections establishing)
tps = 28.323828 (excluding connections establishing)
tps = 27.944931 (excluding connections establishing)
tps = 25.898393 (excluding connections establishing)
tps = 26.727316 (excluding connections establishing)
tps = 27.499692 (excluding connections establishing)
tps = 25.430853 (excluding connections establishing)

with 32k:
> /test/bin/pgbench -c 10 -t 300 test
starting vacuum...end.
transaction type: TPC-B (sort of)
scaling factor: 10
number of clients: 10
number of transactions per client: 300
number of transactions actually processed: 3000/3000
...
tps = 28.609049 (excluding connections establishing)
tps = 29.978503 (excluding connections establishing)
tps = 30.502606 (excluding connections establishing)
tps = 33.406386 (excluding connections establishing)
tps = 30.422134 (excluding connections establishing)
tps = 26.878762 (excluding connections establishing)
tps = 31.461116 (excluding connections establishing)

>>>>> "PJ" == Pailloncy Jean-Gerard <jg(at)rilk(dot)com> writes:

PJ> Hi,
PJ> A small test with 8rc2 and BLCKSZ of 8k and 32k.
PJ> It seems there is a 10% increase in the number of transactions by
PJ> second.
PJ> Does someone plan to carefully test the impact of BLCKSZ ?

One of the suggestions handed to me a long time ago for speeding up PG
on FreeBSD was to double the default blocksize in PG. I tried it, but
found not a significant enough speed up to make it worth the trouble
to remember to patch every version of Pg during the upgrade path (ie,
7.4.0 -> 7.4.2 etc.) Forgetting to do that would be disastrous!

--
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Vivek Khera, Ph.D. Khera Communications, Inc.
Internet: khera(at)kciLink(dot)com Rockville, MD +1-301-869-4449 x806
AIM: vivekkhera Y!: vivek_khera http://www.khera.org/~vivek/

Vivek Khera <khera(at)kcilink(dot)com> writes:
> One of the suggestions handed to me a long time ago for speeding up PG
> on FreeBSD was to double the default blocksize in PG. I tried it, but
> found not a significant enough speed up to make it worth the trouble
> to remember to patch every version of Pg during the upgrade path (ie,
> 7.4.0 -> 7.4.2 etc.) Forgetting to do that would be disastrous!

Not really --- the postmaster will refuse to start if the BLCKSZ shown
in pg_control doesn't match what is compiled in. I concur though that
there may be no significant performance gain. For some workloads there
may well be a performance loss from increasing BLCKSZ.

regards, tom lane

Am Mittwoch, 22. Dezember 2004 22:04 schrieb Tom Lane:
> Vivek Khera <khera(at)kcilink(dot)com> writes:
> > One of the suggestions handed to me a long time ago for speeding up PG
> > on FreeBSD was to double the default blocksize in PG. I tried it, but
> > found not a significant enough speed up to make it worth the trouble
> > to remember to patch every version of Pg during the upgrade path (ie,
> > 7.4.0 -> 7.4.2 etc.) Forgetting to do that would be disastrous!
>
> Not really --- the postmaster will refuse to start if the BLCKSZ shown
> in pg_control doesn't match what is compiled in. I concur though that
> there may be no significant performance gain. For some workloads there
> may well be a performance loss from increasing BLCKSZ.

I've several databases of the same version 7.2 with rowsizes from 8k and 32k
with the same workload (a content management system), and the performance of
the 32k variants is slightly better for a few queries, overall responsivness
seems to better with 8k (maybe because the 8k variant has 4x more buffers).

Regards,
Mario Weilguni