Re: Delaying/avoiding BTreeTupleGetNAtts() call within _bt_compare()

On Thu, Jan 30, 2020 at 1:19 AM Floris Van Nee <florisvannee(at)optiver(dot)com> wrote:
> I'd say applying just v2-0001 is actually slightly hurtful for single-core performance. Applying all of them gives a good improvement though. It looks like the performance improvement is also more noticeable at higher core counts now.

Many thanks for testing once again!

Your tests show that the overall winner is "<v2-0001+2+3>", which is
strictly better than all other configurations you tested -- it is at
least slightly better than every other configuration at every client
count tested. I was particularly pleased to see that "<v2-0001+2+3>"
is ~8.6% faster than the master branch with 30 clients! That result
greatly exceeded my expectations.

I have been able to independently confirm that you really need the
first two patches together to see the benefits -- that wasn't clear
until recently.

The interesting thing now is the role of the "negative infinity test"
patch (the 0003-* patch) in all of this. I suspect that it may not be
helping us much here. I wonder, could you test the following
configurations to settle this question?

* <master> with 30 clients (i.e. repeat the test that you reported on
most recently)

* <v2-0001+2+3> with 30 clients (i.e. repeat the test that you
reported got us that nice ~8.6% increase in TPS)

* <v2-0001+2> with 30 clients -- a new test, to see if performance is
at all helped by the "negative infinity test" patch (the 0003-*
patch).

It seems like a good idea to repeat the other two tests as part of
performing this third test, out of general paranoia. Intel seem to
roll out a microcode update for a spectre-like security issue about
every other day.

Thanks again
--
Peter Geoghegan