Re: Compressed TOAST Slicing

From: Andrey Borodin <x4mmm(at)yandex-team(dot)ru>
To: Paul Ramsey <pramsey(at)cleverelephant(dot)ca>, Владимир Лесков <vladimirlesk(at)yandex-team(dot)ru>
Cc: Stephen Frost <sfrost(at)snowman(dot)net>, Daniel Verite <daniel(at)manitou-mail(dot)org>, Simon Riggs <simon(at)2ndquadrant(dot)com>, rafia(dot)sabih(at)enterprisedb(dot)com, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>, PostgreSQL Hackers <pgsql-hackers(at)lists(dot)postgresql(dot)org>
Subject: Re: Compressed TOAST Slicing
Date: 2019-04-09 17:09:33
Message-ID: C2D8E5D5-3E83-469B-8751-1C7877C2A5F2@yandex-team.ru
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

Hi!

> 12 марта 2019 г., в 10:22, Andrey Borodin <x4mmm(at)yandex-team(dot)ru> написал(а):
>
> 3. And I'd use memmove despite the comment why we do not do that. It is SSE-optimized and cache-optimized nowadays.

So, I've pushed idea a little bit and showed that decompress byte-copy cycle to Vladimir Leskov.
while (len--)
{
*dp = dp[-off];
dp++;
}

He advised me to use algorithm that splits copied regions into smaller non-overlapping subregions with exponentially increasing size.

while (off <= len)
{
memcpy(dp, dp - off, off);
len -= off;
dp += off;
off *= 2;
}
memcpy(dp, dp - off, len);

On original Paul's test without patch of this thread this optimization gave about x2.5 speedup.
I've composed more detailed tests[0] and tested against current master. Now it only gives 20%-25% of decompression speedup, but I think it is still useful.

Best regards, Andrey Borodin.

[0] Here's the test

create table if not exists slicingtest1 as select repeat('0', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest2 as select repeat('01', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest3 as select repeat('012', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest4 as select repeat('0123', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest5 as select repeat('01234', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest6 as select repeat('012345', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest7 as select repeat('0123456', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest8 as select repeat('01234567', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest16 as select repeat('0123456789ABCDEF', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest32 as select repeat('0x1x2x3x4x5x6x7x8x9xAxBxCxDxExFx', 10000) as a from generate_series(1,10000);
create table if not exists slicingtest64 as select repeat('0xyz1xyz2xyz3xyz4xyz5xyz6xyz7xyz8xyz9xyzAxyzBxyzCxyzDxyzExyzFxyz', 10000) as a from generate_series(1,10000);

\timing off
select sum(length(a)) from slicingtest1; -- do for every stride lenght
\timing on
select sum(length(a)) from slicingtest1;

Attachment Content-Type Size
0001-Use-fast-memcpy-in-pglz-decompression.patch application/octet-stream 1.3 KB

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Alvaro Herrera 2019-04-09 17:09:42 Re: Copy function for logical replication slots
Previous Message Fujii Masao 2019-04-09 17:00:03 Re: reloption to prevent VACUUM from truncating empty pages at the end of relation