Re: Patch: Write Amplification Reduction Method (WARM)

On Wed, Apr 12, 2017 at 9:23 AM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com>
wrote:

> On Tue, Apr 11, 2017 at 10:50 PM, Pavan Deolasee
>
> >
> > And I fixed them as quickly as humanly possible.
> >
>
> Yes, you have responded to them quickly, but I didn't get a chance to
> re-verify all of those. However, I think the main point Robert wants
> to say is that somebody needs to dig the complete patch to see if
> there is any kind of problems with it.
>

There are no two views about that. I don't even claim that more problems
won't be found during in-depth review. I was only responding to his view
that I did not do much to address the regressions reported during the
review/tests.

>
> > 5. Added code to set a CLEAR pointer to a WARM pointer when we know that
> the
> > entire chain is WARM. This should address the workload Dilip ran and
> found
> > regression (I don't think he got chance to confirm that)
> >
>
> Have you by any chance tried to reproduce it at your end?

I did reproduce and verified that the new technique helps the case [1] (see
last para). I did not go extra length to check if there are more cases
which can still cause regression, like recheck is applied only once to
each tuple (so the new technique does not yield any benefit) and whether
that still causes regression and by how much. However I ran pure pgbench
workload (only HOT updates) with smallish scale factor so that everything
fits in memory and did not find any regression.

Having said that, it's my view that others need not agree to it, that we
need to distinguish between CPU and IO load since WARM is designed to
address IO problems and not so much CPU problems. We also need to see
things in totality and probably measure updates and selects both if we are
going to WARM update all tuples once and read them once. That doesn't mean
we shouldn't perform more tests and I am more than willing to fix if we
find regression in even a remotely real-world use case.

Thanks,
Pavan

[1]
https://www.postgresql.org/message-id/CABOikdOTstHK2y0rDk%2BY3Wx9HRe%2BbZtj3zuYGU%3DVngneiHo5KQ%40mail.gmail.com

--
Pavan Deolasee http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services