On Mar 29, 2007, at 11:16 AM, Tom Lane wrote:
We've recently made a couple changes to our system that have resulted
in a drastic increase in performance as well as some very confusing
changes to the database statistics, specifically
pg_stat_database.xact_commit. Here's the details:
I'm kinda boggled too. I can see how increasing shared buffers could
result in a drastic reduction in write rate, if the working set of your
queries fits in the new space but didn't fit in the old. I have no idea
how that leads to a drop in number of transactions committed though.
It doesn't make sense that autovac would run less frequently, because
it's driven by number of tuples changed not number of disk writes; and
that could hardly account for a 10x drop anyway.
Did you by any chance take note of exactly which processes were
generating all the I/O or the CPU load?
Well, wrt to the CPU load, as I said, we're pretty sure that's autovac as we still get spikes that hit about the same threshold, after which cache hits go up dramatically and the spikes just don't last two days anymore.
As far as the procs responsible for the writes go, we were unable to see that from the OS level as the guy we had as a systems admin last year totally screwed us with the way he set up the SunCluster on the boxes and we have been unable to run Dtrace which has left us watching a lot of iostat. However, we did notice a direct correlation between write spikes and "write intensive" queries like large COPYs, UPDATEs, and INSERTs.
One very important thing to note here is that the number, or rather rate, of disk writes has not changed. It's the volume of data in those writes that has dropped, along with those transaction mysterious counts. Could the bgwriter be the culprit here? Does anything it does get logged as a transaction?
software developer
615-296-0838
emma(r)