From: | Neil Conway <neil(dot)conway(at)gmail(dot)com> |
---|---|
To: | PostgreSQL Hackers <pgsql-hackers(at)postgresql(dot)org> |
Subject: | Optimizing COPY with SIMD |
Date: | 2024-06-02 19:17:21 |
Message-ID: | CAOW5sYb1HprQKrzjCsrCP1EauQzZy+njZ-AwBbOUMoGJHJS7Sw@mail.gmail.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
Inspired by David Rowley's work [1] on optimizing JSON escape processing
with SIMD, I noticed that the COPY code could potentially benefit from SIMD
instructions in a few places, eg:
(1) CopyAttributeOutCSV() has 2 byte-by-byte loops
(2) CopyAttributeOutText() has 1
(3) CopyReadLineText() has 1
(4) CopyReadAttributesCSV() has 1
(5) CopyReadAttributesText() has 1
Attached is a quick POC patch that uses SIMD instructions for case (1)
above. For sufficiently large attribute values, this is a significant
performance win. For small fields, performance looks to be about the same.
Results on an M1 Macbook Pro.
======
neilconway=# select count(*), avg(length(a))::int, avg(length(b))::int,
avg(length(c))::int from short_strings;
count | avg | avg | avg
--------+-----+-----+-----
524288 | 8 | 8 | 8
(1 row)
neilconway=# select count(*), avg(length(a))::int, avg(length(b))::int,
avg(length(c))::int from long_strings;
count | avg | avg | avg
-------+-----+-----+-----
65536 | 657 | 657 | 657
(1 row)
master @ 8fea1bd541:
$ for i in ~/*.sql; do hyperfine --warmup 5 "./psql -f $i"; done
Benchmark 1: ./psql -f /Users/neilconway/copy-out-bench-long-quotes.sql
Time (mean ± σ): 2.027 s ± 0.075 s [User: 0.001 s, System: 0.000
s]
Range (min … max): 1.928 s … 2.207 s 10 runs
Benchmark 1: ./psql -f /Users/neilconway/copy-out-bench-long.sql
Time (mean ± σ): 1.420 s ± 0.027 s [User: 0.001 s, System: 0.000
s]
Range (min … max): 1.379 s … 1.473 s 10 runs
Benchmark 1: ./psql -f /Users/neilconway/copy-out-bench-short.sql
Time (mean ± σ): 546.0 ms ± 9.6 ms [User: 1.4 ms, System: 0.3 ms]
Range (min … max): 539.0 ms … 572.1 ms 10 runs
master + SIMD patch:
$ for i in ~/*.sql; do hyperfine --warmup 5 "./psql -f $i"; done
Benchmark 1: ./psql -f /Users/neilconway/copy-out-bench-long-quotes.sql
Time (mean ± σ): 797.8 ms ± 19.4 ms [User: 0.9 ms, System: 0.0 ms]
Range (min … max): 770.0 ms … 828.5 ms 10 runs
Benchmark 1: ./psql -f /Users/neilconway/copy-out-bench-long.sql
Time (mean ± σ): 732.3 ms ± 20.8 ms [User: 1.2 ms, System: 0.0 ms]
Range (min … max): 701.1 ms … 763.5 ms 10 runs
Benchmark 1: ./psql -f /Users/neilconway/copy-out-bench-short.sql
Time (mean ± σ): 545.7 ms ± 13.5 ms [User: 1.3 ms, System: 0.1 ms]
Range (min … max): 533.6 ms … 580.2 ms 10 runs
======
Implementation-wise, it seems complex to use SIMD when
encoding_embeds_ascii is true (which should be uncommon). In principle, we
could probably still use SIMD here, but it would require juggling between
the SIMD chunk size and sizes returned by pg_encoding_mblen(). For now, the
POC patch falls back to the old code path when encoding_embeds_ascii is
true.
Any feedback would be very welcome.
Cheers,
Neil
Attachment | Content-Type | Size |
---|---|---|
0002-Optimize-COPY-TO-.-FORMAT-CSV-using-SIMD-instruction.patch | application/octet-stream | 5.8 KB |
0001-Remove-inaccurate-comment.patch | application/octet-stream | 738 bytes |
copy-out-bench-short.sql | application/octet-stream | 375 bytes |
copy-out-bench-long.sql | application/octet-stream | 373 bytes |
copy-out-bench-long-quotes.sql | application/octet-stream | 403 bytes |
From | Date | Subject | |
---|---|---|---|
Next Message | Andrew Dunstan | 2024-06-02 19:47:29 | Re: meson and check-tests |
Previous Message | Andrew Dunstan | 2024-06-02 18:39:02 | Re: The xversion-upgrade test fails to stop server |