Jeff Janes wrote:
> In my test cases, the syncs that the backends were doing were almost
> always to the same file that the checkpoint writer was already choking
> on (so they are entangled simply by virtue of that). So very quickly
> all the backends hit the same wall and thunked to a halt. This is
> probably a feature of trying to use pgbench as the basis to get a very
> artificial model.
Yes--pgbench has some problems like you describe, ones that are a bit
different than the way I've seen fsync writes get in each other's way in
the production systems I've looked at. That's good if you really want
to provoke this behavior, which is one reason why I've used as an
example for my patches so far (the other being that it's already
available in everyone's installation). But it's tough to get it to act
more like a real-world system, which don't have quite so many localized
updates, without cranking the scale way up. And that then tends to
aggravate other problems too.
The 8.3 checkpoint spreading work also got some useful results using the
dbt-2 benchmark. I'm at the point where I think I need to return to
that test program for what I'm doing now. I'd encourage you to try that
out too if you get a chance.
Thanks for the feedback and the review. I hope you appreciate now why I
suggested you wait for the stuff I was submitting before getting back
into the sorted checkpoint topic again. That should be a lot easier to
make sense of with this instrumentation in place.
--
Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD
PostgreSQL Training, Services and Support www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books