Обсуждение: More on 7.4b2 vs 7.3.4 performance

Per Tom Lane's suggestion, I increased the pgbench
scale factor to be equal to the max number of clients
in each test run.

Attached sample graph shows 1 - 10 clients,
100-500 transactions per client, with scale
factor=10 for all runs.

There is still a average 9% improvement in 7.4b2
transaction rates over 7.3.4.  For low number
of clients (1 - 3) performance was almost the
same.  For higher numbers of clients, the 7.4
advantage was often in the 15-25% range.  Other
tests with scale factor == number of clients and
other pertubations gave similar results.
7.4 is almost always faster than 7.3.4.

Mind you now, I am *not* complaining.

Josh Rovero wrote:
> Per Tom Lane's suggestion, I increased the pgbench
> scale factor to be equal to the max number of clients
> in each test run.
>
> Attached sample graph shows 1 - 10 clients,
> 100-500 transactions per client, with scale
> factor=10 for all runs.
>
> There is still a average 9% improvement in 7.4b2
> transaction rates over 7.3.4.  For low number
> of clients (1 - 3) performance was almost the
> same.  For higher numbers of clients, the 7.4
> advantage was often in the 15-25% range.  Other
> tests with scale factor == number of clients and
> other pertubations gave similar results.
> 7.4 is almost always faster than 7.3.4.
>
> Mind you now, I am *not* complaining.

That's closer to the performance improvement we were expecting.

--
  Bruce Momjian                        |  http://candle.pha.pa.us
  pgman@candle.pha.pa.us               |  (610) 359-1001
  +  If your life is a hard drive,     |  13 Roberts Road
  +  Christ can be your backup.        |  Newtown Square, Pennsylvania 19073

Josh Rovero <rovero@sonalysts.com> writes:
> There is still a average 9% improvement in 7.4b2
> transaction rates over 7.3.4.  For low number
> of clients (1 - 3) performance was almost the
> same.  For higher numbers of clients, the 7.4
> advantage was often in the 15-25% range.

This seems more believable as a version-to-version improvement factor.

I suspect that your previous numbers reflect some isolated tweak that
we made in the lock management code, that happened to reduce the amount
of time wasted in a heavy-contention scenario.  Not sure what though...

            regards, tom lane