Re: Yet another vectorized engine - Mailing list pgsql-hackers

Hi Konstantin,

I checkout your branch pg13 in repo https://github.com/zhangh43/vectorize_engine

After I fixed some compile error, I tested Q1 on TPCH-10G

The result is different from yours and vectorize version is too slow. Note that I disable parallel worker by default.

no JIT no Vectorize:  36 secs

with JIT only:             23 secs

with Vectorize only:   33 secs

JIT + Vectorize:         29 secs

My config option is `CFLAGS='-O3 -g -march=native' --prefix=/usr/local/pgsql/ --disable-cassert --enable-debug --with-llvm`

I will do some spike on why vectorized is so slow. Could you please provide your compile option and the TPCH dataset size and your queries(standard Q1?) to help me to debug on it.

Hi, Hubert

Sorry, looks like I have used slightly deteriorated snapshot of master so I have not noticed some problems.
Fixes are committed.

Most of the time is spent in unpacking heap tuple (tts_buffer_heap_getsomeattrs):

  24.66%  postgres  postgres             [.] tts_buffer_heap_getsomeattrs
   8.28%  postgres  vectorize_engine.so  [.] VExecStoreColumns
   5.94%  postgres  postgres             [.] HeapTupleSatisfiesVisibility
   4.21%  postgres  postgres             [.] bpchareq
   4.12%  postgres  vectorize_engine.so  [.] vfloat8_accum


In my version of nodeSeqscan I do not keep all fetched 1024 heap tuples but stored there attribute values in vector columns immediately.
But to avoid extraction of useless data it is necessary to know list of used columns.
The same problem is solved in zedstore, but unfortunately there is no existed method in Postgres to get list
of used attributes. I have done it but my last implementation contains error which cause loading of all columns.
Fixed version is committed.

Now profile without JIT is:

 15.52%  postgres  postgres             [.] tts_buffer_heap_getsomeattrs
  10.25%  postgres  postgres             [.] ExecInterpExpr
   6.54%  postgres  postgres             [.] HeapTupleSatisfiesVisibility
   5.12%  postgres  vectorize_engine.so  [.] VExecStoreColumns
   4.86%  postgres  postgres             [.] bpchareq
   4.80%  postgres  vectorize_engine.so  [.] vfloat8_accum
   3.78%  postgres  postgres             [.] tts_minimal_getsomeattrs
   3.66%  postgres  vectorize_engine.so  [.] VExecAgg
   3.38%  postgres  postgres             [.] hashbpchar

and with JIT:

 13.88%  postgres  postgres             [.] tts_buffer_heap_getsomeattrs
   7.15%  postgres  vectorize_engine.so  [.] vfloat8_accum
   6.03%  postgres  postgres             [.] HeapTupleSatisfiesVisibility
   5.55%  postgres  postgres             [.] bpchareq
   4.42%  postgres  vectorize_engine.so  [.] VExecStoreColumns
   4.19%  postgres  postgres             [.] hashbpchar
   4.09%  postgres  vectorize_engine.so  [.] vfloat8pl

On Mon, Feb 24, 2020 at 8:43 PM Hubert Zhang <hzhang@pivotal.io> wrote:

Hi Konstantin,

I have added you as a collaborator on github. Please accepted and try again.

I think non collaborator could also open pull requests.

On Mon, Feb 24, 2020 at 8:02 PM Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote:

On 24.02.2020 05:08, Hubert Zhang wrote:

Hi

On Sat, Feb 22, 2020 at 12:58 AM Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote:

On 12.02.2020 13:12, Hubert Zhang wrote:

On Tue, Feb 11, 2020 at 1:20 AM Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote:

So looks like PG-13 provides significant advantages in OLAP queries comparing with 9.6!
Definitely it doesn't mean that vectorized executor is not needed for new version of Postgres.
Once been ported, I expect that it should provide comparable  improvement of performance.

But in any case I think that vectorized executor makes sense only been combine with columnar store.

Thanks for the test. +1 on vectorize should be combine with columnar store. I think when we support this extension

on master, we could try the new zedstore.

I'm not active on this work now, but will continue when I have time. Feel free to join bring vops's feature into this extension.

 

Thanks

Hubert Zhang

I have ported vectorize_engine to the master.
It takes longer than I expected: a lot of things were changed in executor.

Results are the following:

par.warkers	PG9_6 vectorize=off	PG9_6 vectorize=on	master vectorize=off jit=on	master vectorize=off jit=off	master vectorize=on jit=ofn	master vectorize=on jit=off
0	36	20	16	25.5	15	17.5
4	10	-	5	7	-	-

So it proves the theory that JIT provides almost the same speedup as vector executor (both eliminates interpretation overhead but in different way).
I still not sure that we need vectorized executor: because with standard heap it provides almost no improvements comparing with current JIT version.
But in any case I am going to test it with vertical storage (zedstore or cstore).

Thanks for the porting and testing.

Yes, PG master and 9.6 have many changes, not only executor, but also tupletableslot interface.

What matters the performance of JIT and Vectorization is its implementation. This is just the beginning of vectorization work, just as your vops extension reported, vectorization could run 10 times faster in PG. With the overhead of row storage(heap), we may not reach that speedup, but I think we could do better. Also +1 on vertical storage.

BTW, welcome to submit your PR for the PG master version.

Sorry, but I have no permissions to push changes to your repository.
I can certainly create my own fork of vectorize_engine, but I think it will be beter if I push pg13 branch in your repository.

--

Thanks

Hubert Zhang

--

Thanks

Hubert Zhang

-- 
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company