Checkpoint cost, looks like it is WAL/CRC

Tom, All:

Ok, finally managed though the peristent efforts of Mark Wong to get some
tests through. Here are two tests with the CRC and wall buffer checking
completely cut out of the code, as Tom suggested:

5-min checkpoint:
http://khack.osdl.org/stp/302738/results/0/
http://khack.osdl.org/stp/302706/results/0/
60-min checkpoint:
http://khack.osdl.org/stp/302739/results/0/
(please note that OSDL is having technical difficulties and some links may
not work)

This is the performance profile I'd expect and want to see, and the
frequency for checkpoints doesn't affect the overall performance at all.
Contrast it with these:

5-min checkpoint:
http://khack.osdl.org/stp/302671/results/0/
I don't hae a 60-minute checkpoint for comparison because of failures on
the STP :-(

So, now that we know what the performance bottleneck is, how do we fix it?

--
--Josh

Josh Berkus
Aglio Database Solutions
San Francisco