Greg Smith, 25.02.2010 17:47:
> Based on tests showing a similar style and magnitude regression at Sun
> by Jignesh Shah, I would assume this is mainly because some of the
> starting parameter changes in 8.4 detuned this particular benchmark a
> bit, in favor of proving a better default for real-world users. For
> example, the starting default_statistics_target was raised from 10 to
> 100 in 8.4. This causes a mild decrease in performance on trivial
> benchmarks like this one, while potentially providing a large
> improvement in the sorts of query plans seen in real applications.
>
> That was the basic theme for the sorts of performance changes that
> showed up in 8.4. Another example (not actually relevant to this
> benchmark) is that the Free Space Map used to track deleted items is now
> kept on disk instead of in shared memory. That's obviously less
> efficient in the short term--disk write instead of just a memory
> one--but it prevents all sorts of nasty worst-case scenarios you used to
> run into the FSM wasn't big enough in earlier versions. Basically, the
> 8.4 performance related changes reduced average performance on trivial
> benchmark workloads a small amount, in favor of large improvements in
> the sort of situations people run into in production deployments. I
> think it was the right trade-off to make.
Thanks for the detailed answer!
Regards
Thomas