Hi,
On 2020-06-12 14:37:15 -0700, Andres Freund wrote:
> On 2020-06-11 11:14:02 -0700, Jeff Davis wrote:
> > On Thu, 2020-06-11 at 10:45 -0700, Andres Freund wrote:
> > > Did you run any performance tests?
> >
> > Yes, I reproduced your ~12% regression from V12, and this patch nearly
> > eliminated it for me.
>
> I spent a fair bit of time looking at the difference. Jeff had let me
> know on chat that he was still seeing some difference, but couldn't
> quite figure out where that was.
>
> Trying it out myself, I observed that the patch helped, but not that
> much. After a bit I found one major reason for why:
> LookupTupleHashEntryHash() assigned the hash to pointer provided by the
> caller's before doing the insertion. That ended up causing a pipeline
> stall (I assume it's store forwarding, but not sure). Moving the
> assignment to the caller variable to after the insertion got rid of
> that.
This is still not resolved. We're right now slower than 12. It's
effectively not possible to do performance comparisons right now. This
was nearly two months ago.
Greetings,
Andres Freund