Re: POC: Lock updated tuples in tuple_update() and tuple_delete()

On Thu, Mar 2, 2023 at 11:16 PM Alexander Korotkov <aekorotkov(at)gmail(dot)com> wrote:
> On Thu, Mar 2, 2023 at 9:17 PM Pavel Borisov <pashkin(dot)elfe(at)gmail(dot)com> wrote:
> > On Thu, 2 Mar 2023 at 18:53, Alexander Korotkov <aekorotkov(at)gmail(dot)com> wrote:
> > > On Thu, Mar 2, 2023 at 1:29 PM Pavel Borisov <pashkin(dot)elfe(at)gmail(dot)com> wrote:
> > > > > Let's see the performance results for the patchset. I'll properly
> > > > > revise the comments if results will be good.
> > > > >
> > > > > Pavel, could you please re-run your tests over revised patchset?
> > > >
> > > > Since last time I've improved the test to avoid significant series
> > > > differences due to AWS storage access variation that is seen in [1].
> > > > I.e. each series of tests is run on a tmpfs with newly inited pgbench
> > > > tables and vacuum. Also, I've added a test for low-concurrency updates
> > > > where the locking optimization isn't expected to improve performance,
> > > > just to make sure the patches don't make things worse.
> > > >
> > > > The tests are as follows:
> > > > 1. Heap updates with high tuple concurrency:
> > > > Prepare without pkeys (pgbench -d postgres -i -I dtGv -s 10 --unlogged-tables)
> > > > Update tellers 100 rows, 50 conns ( pgbench postgres -f
> > > > ./update-only-tellers.sql -s 10 -P10 -M prepared -T 600 -j 5 -c 50 )
> > > >
> > > > Result: Average of 5 series with patches (0001+0002) is around 5%
> > > > faster than both master and patch 0001. Still, there are some
> > > > fluctuations between different series of the measurements of the same
> > > > patch, but much less than in [1]
> > >
> > > Thank you for running this that fast!
> > >
> > > So, it appears that 0001 patch has no effect. So, we probably should
> > > consider to drop 0001 patch and consider just 0002 patch.
> > >
> > > The attached patch v12 contains v11 0002 patch extracted separately.
> > > Please, add it to the performance comparison. Thanks.
> >
> > I've done a benchmarking on a full series of four variants: master vs
> > v11-0001 vs v11-0001+0002 vs v12 in the same configuration as in the
> > previous measurement. The results are as follows:
> >
> > 1. Heap updates with high tuple concurrency:
> > Average of 5 series v11-0001+0002 is around 7% faster than the master.
> > I need to note that while v11-0001+0002 shows consistent performance
> > improvement over the master, its value can not be determined more
> > precisely than a couple of percents even with averaging. So I'd
> > suppose we may not conclude from the results if a more subtle
> > difference between v11-0001+0002 vs v12 (and master vs v11-0001)
> > really exists.
> >
> > 2. Heap updates with high tuple concurrency:
> > All patches and master are still the same within a tolerance of
> > less than 0.7%.
> >
> > Overall patch v11-0001+0002 doesn't show performance degradation so I
> > don't see why to apply only patch 0002 skipping 0001.
>
> Thank you, Pavel. So, it seems that we have substantial benefit only
> with two patches. So, I'll continue working on both of them.

The revised patchset is attached. The patch removing extra
table_tuple_fetch_row_version() is back. The second patch now
implements a concept of LazyTupleTableSlot, a slot which gets
allocated only when needed. Also, there is more minor refactoring and
more comments.

------
Regards,
Alexander Korotkov