Newer
Older
* COMMENT -*- mode: org -*-
#+Date: 2016-12-16
* database binding benchmark
I want to know the efficiency of the db binding for different language,
libraries and database servers.
- table:
CREATE TABLE foo
(id integer primary key,
bar integer);
INSERT INTO foo (id, bar) VALUES (1, 1);
- run many times:
SELECT * FROM foo
UPDATE foo SET bar=bar+1 WHERE id=1
* benchmark results on x201 laptop, postgres runs on 5400rpm SATA disk.
each entry has format: db_server, compiler, libraries
all build is release mode (with -O or equivalent).
| db | language | runtime | lib | select (s) | update (s) |
|-------+-------------+---------------+-------------------+------------+------------|
| pg9.4 | python | cpython 2.7.9 | psycopg2 | 0.37 | 26.3 |
| pg9.4 | C | gcc 4.9.2 | libpq5 | 0.32 | 26.5 |
| pg9.4 | haskell | ghc 8.0.1 | postgresql-simple | 0.63 | 27.2 |
| pg9.4 | haskell | ghc 7.10.3 | postgresql-simple | 0.61 | 26.6 |
| pg9.4 | rust | rust 1.13.0 | postgres | 0.50 | 26.3 |
| pg9.4 | go | go 1.7.3 | pq | 0.37 | 27.2 |
| pg9.4 | racket | racket 6.5 | built-in db | 1.99 | 26.7 |
| pg9.4 | common lisp | sbcl 1.2.11 | postmodern | 0.57 | 27.0 |
- Remarks:
- python and go's db binding is as fast as C.
- rust, haskell, cl is slower. C is 1.5x faster.
** postgres 9.4, cpython 2.7, psycopg2==2.6.1
http://initd.org/psycopg/docs/usage.html
0.35
28.77
--
0.38
26.15
--
0.37
26.35
** postgres 9.4, C, libpq5
https://www.postgresql.org/docs/9.4/static/libpq.html
0.33
26.6
--
0.32
26.5
--
0.30
26.2
//side note: why adding -shared in LDFLAGS will just segfault the program?
** postgres 9.4, lts-7.13, ghc 8.0.1, postgresql-simple
0.71
27.5
--
0.63
27.2
--
0.61
26.2
** postgres 9.4, lts-5.18, ghc 7.10.3, postgresql-simple
0.61
26.6
--
0.59
26.8
--
0.63
26.6
** postgres 9.4, rust 1.13.0, postgres 0.13
0.48
26.3
--
0.53
26.2
--
0.50
28.0
** postgres 9.4, go 1.7.3, pq
0.385
27.0
--
0.367
27.6
--
0.423
30.7
--
0.362
27.2
--
0.372
26.4
** postgres 9.4, racket 6.5
2.21
26.72
--
1.85
26.63
- I think I maybe testing disk IO performance. should use the pg bench cluster
for benchmarking. bench cluster listen on port 5433.
when running against bench cluster.
result is
2.15
2.29
--
2.10
2.21
So write performance is definitely testing disk IO performance. Read
performance is related to the language and runtime.
- try compile it with raco exe first.
2.00
27.84
--
2.02
27.62
it's about the same.
- try use query-rows instead of query.
1.78
26.94
--
2.02
29.08
--
1.99
26.79
it's still the slowest.
** postgres 9.4, sbcl 1.2.11, postmodern 20160208
0.54
26.69
--
0.57
27.02
--
0.58
27.16
- try enable optimize
C-c C-k with negative prefix
time arithmetic doesn't matter.
0.58
27.28
--
0.54
28.51
no effect.
* benchmark result on new ryzen5 PC. postgres runs on M2 SSD.
** 2019-02-23 pg 9.6, python 2.7.13, psycopg2==2.6.2
INFO running benchmark...
DEBUG trace_time, func=select_test, duration=0.18
DEBUG trace_time, func=update_test, duration=2.78
--
INFO running benchmark...
DEBUG trace_time, func=select_test, duration=0.19
DEBUG trace_time, func=update_test, duration=2.76
** 2019-02-23 pg 9.6. openjdk 1.8.0_162, jdbc postgresql 42.2.5
warming up JVM...
select time cost: 0.29s
update time cost: 2.68s
--
warming up JVM...
select time cost: 0.27s
update time cost: 2.67s
--
select time cost: 0.35s
update time cost: 2.72s
--
select time cost: 0.33s
update time cost: 2.75s
--
** 2019-02-23 pg 9.6, ghc 8.2.2, lts 11.7
seconds: 0.249821547
seconds: 2.703614265
--
seconds: 0.252873544
seconds: 2.662445433
--
seconds: 0.244120763
seconds: 2.655961013
--
** 2019-02-23 pg 9.6, gcc 6.3.0, libpq-dev 10.3-1
0.170
2.598
--
0.172
2.551
--
0.178
2.562
--
0.172
2.589