On 12/14/06, Tom Lane <tgl@xxxxxxxxxxxxx> wrote:
Evgeny Gridasov <eugrid@xxxxxxxxxxxx> writes:
> This is a Linux Debian 3.1 ontop of 2x XEON 3.4 Ghz.
> PostgreSQL is 8.2 checked out from CVS REL8_2_STABLE yesterday.
> I'm running the same Postgres on another machine,
> with Debian Etch and have the same results.
Hmph. With 8.2 on Fedora 5 on a 2.8Ghz dual Xeon, I get this:
<snip>
regression=# explain analyze select count(*) from foo;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------
Aggregate (cost=44764.00..44764.01 rows=1 width=0) (actual time=1324.846..1324.847 rows=1 loops=1)
-> Seq Scan on foo (cost=0.00..38514.00 rows=2500000 width=0) (actual time=0.046..748.582 rows=2500000 loops=1)
Total runtime: 1324.902 ms
(3 rows)
Time: 1325.591 ms
regression=#
which works out to about 0.14 microsec per gettimeofday call, on a
machine that ought to be slower than yours. So I think you've got
either a crummy motherboard, or a kernel that doesn't know the best
way to read the clock on that hardware. There is some discussion
of this in the archives (probably in pgsql-hackers); look back around
May or so when we were unsuccessfully trying to hack EXPLAIN to use
fewer gettimeofday calls.
Yow! I notice the same thing on our HP BL25p blades w/2*opteron 270
(four total cores, AMD 8111 or 8131 chipset). 1.25 microsec/call vs
my new desktop (Intel Core2 6300) 0.16 microsec/call.
I hope this isn't a "crummy mainboard" but I can't seem to affect
things by changing clock source (kernel 2.6.16 SLES10). I tried
kernel command option clock=XXX where XXX in
(cyclone,hpet,pmtmr,tsc,pit), no option was significantly better than
the default.
Anyone know how this might be improved (short of replacing hardware)?
-K
