On Mon, 2004-12-06 at 17:42, Mark Wong wrote:
> On Tue, Nov 30, 2004 at 10:51:42PM +0000, Simon Riggs wrote:
> > My suggestion: increase checkpoint_timeout to 600 secs, increase
> > bgwriter parameters also, to reduce how frequently it is called, as well
> > as increase the number of blocks per cycle.
>
> Ok, here are a series of three tests varying the bgwriter_delay at 1,
> 50, and 100:
> http://www.osdl.org/projects/dbt2dev/results/pgsql/bgwriter_delay/
>
> I also reduced checkpoint_timeout to 600, where it was prevously at
> 1800 with the results I posted previously. The throughput changes
> weren't significant and the oprofile data is more of less the same.
> I'll try varying the bgwriter_maxpages next (that's the number of
> blocks per cycle, right?)
Mark,
Few questions:
- can we put the logging to DEBUG1 please, so we can see the
checkpoints? ...and set debug_shared_buffers = 10
I don't understand why the checkpoints are so regular at 300 seconds if
the checkpoint_timeout is set to 1800 or other...exactly when and how
are those parameters provided to the server?
- can we set checkpoint_segments to 8192 just to see if that changes the
checkpoint frequency (it should)
- the log output shows the database starts about 4 hours before the main
test starts... err whats going on there? maybe we could get more tests
in if that time could be reduced
- the explain plan output is missing...
http://www.osdl.org/projects/dbt2dev/results/dev4-010/199/db/plan0.out.gz
- the log output shows deadlocks occurring - is there something about
the application of DBT-2 which is actually causing contention? Might
that have changed between beta4 and beta5? The earlier lock waits we saw
("Exclusive Lock" thread) are likely to be related to that. Is there
some other artifact of the test that could cause this...random number
generator....etc. My understanding was that TPC-C didn't deadlock, but I
could be wrong there. This could easily be throwing off the test
results... usually to do with the order in which locks are occurring...
if its not, I hope its not a bug,,,
--
Best Regards, Simon Riggs
