On Wed, 20 Jun 2007, Mike Benoit wrote:
> It would be cool if someone started a generic configuration+benchmark
> utility that could be used with virtually any software.
It would be cool. It would also be impossible.
> Create a benchmark script for the application that returns relevant
> metrics. In PGSQL's case, it would be tied in to PG bench probably. In
> Apache's case AB. This utility would of course need to know how to read
> the metrics to determine what is "best".
The usual situation in these benchmarks is that you get parameters that
adjust along a curve where there's a trade-off between, say, total
throughput and worse-case latency. Specifying "best" here would require a
whole specification language if you want to model how real tuning efforts
work. The AB case is a little simpler, but for PostgreSQL you'd want
something like "With this database and memory sizing, I want the best
throughput possible where maximum latency is usually <5 seconds with 1-30
clients running this transaction, while still maintaining at least 400 TPS
with up to 100 clients, and the crash recovery time can't take more than
10 minutes". There are all sorts of local min/max situations and
non-robust configurations an automated tool will put you into if you don't
force an exhaustive search by being very specific like this.
> I don't think something like this would be very difficult at all to
> write
Here I just smile and say that proves you've never tried to write one :)
It's a really hard problem that gets harder the more you poke at it.
There's certainly lots of value to writing a utility that automatically
tests out multiple parameter values in a batch and compares the results.
If you're not doing that now, you should consider scripting something up
that does. Going beyond that to having it pick the optimal parameters
more automatically would take AI much stronger than just a genetic
algorithm approach.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD