On Wednesday, June 27, 2018 5:38:31 PM CEST Pavel Stehule wrote:
> 2018-06-27 17:19 GMT+02:00 Tomas Vondra <tomas.vondra@2ndquadrant.com>:
> > On 06/26/2018 09:25 PM, Pavel Stehule wrote:
> >> Hi
> >>
> >> ...
> >>
> >> So I am able to see effect of jit_tuple_deforming, and very well, but
> >> only if optimization is active. When optimization is not active then
> >> jit_tuple_deforming does slowdown.
> >>
> >> So maybe a usage of jit_tuple_deforming can be conditioned by
> >> jit_optimization?
> >
> > Can you share the test case and some detail about the hardware and
> > PostgreSQL configuration?
>
> I did very simple test
>
>
> 0.
>
> master branch without asserts, shared buffer to 1GB
> tested on Lenovo T520 8GB RAM 8CPU, i7
> Fedora 28, gcc CFLAGS="-ggdb -Og -g3 -fno-omit-frame-pointer" --with-llvm
>
> 1.
>
> select 'create table wt(' || string_agg(format('%I int', 'c' || i),',') ||
> ')' from generate_series(1,100) g(i) \gexec
>
>
> 2.
>
> begin;
> select 'insert into wt values(' || (select
> string_agg((random()*10000)::int::text,',') from generate_series(1,j - j +
> 100) g(i)) || ')' from generate_series(1,1000000) gg(j) \gexec
> insert into wt select * from wt;
> commit;
>
> 3.
>
> set max_paralel_workers to 0; -- the effect of JIT will be more visible
>
> analyze wt;
> \timing
>
> select sum(c99) from wt;
>
> I tested some combination of:
>
> jit: off on
> jit_inline_above_cost: 0, 10000000000000
> jit_optimize_above_cost: 0, 10000000000000
> jit_tuple_deforming: on, off
>
>
> My primitive tests shows nice possitive effect of jit_tuple_deforming if
> jit optimization is active. When jit optimization is not active, then
> jit_tuple_deforming did slowdown in my test.
>
> So there is range of costs between 100000 and 500000 where
> jit_tuple_deforming didn't work well (without optimization)
>
> I am limmited by small memory of my notebook - when I created table larger
> than 3GB, then I got IO waits on my crypted disc, and any effect of JIT was
> eliminated.
>
> Regards
>
> Pavel
Hi
I have studied this case a bit, and I think too that there is something wrong
here.
Right now, jit_optimize is a -O3. It's really expensive, and triggering it
here is not the right solution. In the attached patch, I force a -O1 for tuple
deforming. With your test case, on my computer, the results are :
- no jit : 850ms
- jit with tuple deforming without optimizations : 1650 ms (1.5ms spent
optimizing)
- jit without tuple deforming : 820ms (0.2ms)
- jit with tuple deforming with optimization (-O3) : 770ms (105ms)
- jit with tuple deforming with patch (-O1) : 725ms (54ms)
I will look at the generated code for tuple deforming, but I think we should
pre-optimize the LLVM bytecode if we do not want to involve the LLVM
optimization passes.
Regards
Pierre
