q4m の configure のオプションによる速度の違い
q4m の configure のオプション変更によってどの程度パフォーマンスに違いが出るのか比較してみた。
Intel(R) Xeon(TM) CPU 2.80GHz x 4 CPU
OS : Debian(Etch)
まずは、default 状態。
./configure --with-mysql=/home/kameid/mysql-5.1.30/ --prefix=/usr/local/mysql
暗黙的に、以下のオプションを指定していることになる。
--with-sync=yes commit to disk at checkpoints (default)
--with-delete=pwrite use pwrite for row deletions (default)
t/05-multireader..........................ok 1/4
Multireader benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 3.991 seconds
Throughput: 1603.607 mess./sec.
t/05-multireader..........................ok
t/05-multirw..............................ok 1/4
Multi-reader-writer benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 10.419 seconds
Throughput: 614.243 mess./sec.
t/05-multirw..............................ok
t/05-multiwait............................ok 1/4
Multi-reader-writer benchmark result under semi-starvation:
Number of messages: 6400
Number of readers: 32
Elapsed: 19.429 seconds
Throughput: 329.396 mess./sec.
with-sync=no を指定。
./configure --with-mysql=/home/ts-dkamei/mysql-5.1.30/ --prefix=/usr/local/mysql --with-sync=no
t/05-multireader..........................ok 1/4
Multireader benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 1.451 seconds
Throughput: 4409.365 mess./sec.
t/05-multireader..........................ok
t/05-multirw..............................ok 1/4
Multi-reader-writer benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 1.620 seconds
Throughput: 3950.522 mess./sec.
t/05-multirw..............................ok
t/05-multiwait............................ok 1/4
Multi-reader-writer benchmark result under semi-starvation:
Number of messages: 6400
Number of readers: 32
Elapsed: 2.544 seconds
Throughput: 2515.237 mess./sec.
delete=msync を指定。
./configure --with-mysql=/home/ts-dkamei/mysql-5.1.30/ --prefix=/usr/local/mysql --with-delete=msync
t/05-multireader..........................ok 1/4
Multireader benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 2.897 seconds
Throughput: 2208.953 mess./sec.
t/05-multireader..........................ok
t/05-multirw..............................ok 1/4
Multi-reader-writer benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 11.628 seconds
Throughput: 550.380 mess./sec.
t/05-multirw..............................ok
t/05-multiwait............................ok 1/4
Multi-reader-writer benchmark result under semi-starvation:
Number of messages: 6400
Number of readers: 32
Elapsed: 19.676 seconds
Throughput: 325.266 mess./sec.
./configure --with-mysql=/home/ts-dkamei/mysql-5.1.30/ --prefix=/usr/local/mysql --with-sync=no --with-delete=msync
t/05-multireader..........................ok 1/4
Multireader benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 1.371 seconds
Throughput: 4669.161 mess./sec.
t/05-multireader..........................ok
t/05-multirw..............................ok 1/4
Multi-reader-writer benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 1.410 seconds
Throughput: 4540.192 mess./sec.
t/05-multirw..............................ok
t/05-multiwait............................ok 1/4
Multi-reader-writer benchmark result under semi-starvation:
Number of messages: 6400
Number of readers: 32
Elapsed: 2.486 seconds
Throughput: 2574.933 mess./sec.
./configure --with-mysql=/home/ts-dkamei/mysql-5.1.30/ --prefix=/usr/local/mysql --enable-mmap-writes=yes
t/05-multireader..........................ok 1/4
Multireader benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 4.008 seconds
Throughput: 1596.810 mess./sec.
t/05-multireader..........................ok
t/05-multirw..............................ok 1/4
Multi-reader-writer benchmark result:
Number of messages: 6400
Number of readers: 32
Elapsed: 12.476 seconds
Throughput: 512.998 mess./sec.
t/05-multirw..............................ok
t/05-multiwait............................ok 1/4
Multi-reader-writer benchmark result under semi-starvation:
Number of messages: 6400
Number of readers: 32
Elapsed: 18.725 seconds
Throughput: 341.785 mess./sec.
| Multireader | multirw | multiwait | |
| with-sync=yes,with-delete=pwrite | 1603 | 614 | 329 |
| with-sync=no,with-delete=pwrite | 4409 | 3950 | 2515 |
| with-sync=yes,with-delete=msync | 2208 | 550 | 325 |
| with-sync=no,with-delete=msync | 4669 | 4540 | 2574 |
| enable-mmap-writes=yes | 1596 | 512 | 341 |
この結果によると、とりあえず、
./configure --with-mysql=/home/ts-dkamei/mysql-5.1.30/ --prefix=/usr/local/mysql --with-sync=no --with-delete=msync
にしとけ、という感じになりそうだが、with-sync=no にすることで、コミット時にHDDに物理的にデータが書き込みされるとは限らなくなるため、大幅に速度がUPしているのである。障害発生時のことを考えると、微妙なところだろう。
delete 時 pwrite か msync か、というのはそれほど影響は無い様だが、とりあえず、msync にしとけばいいだろという数字。enable-mmap-writes=yes は意味が無いことの確認。32bit 版の OS を使用しているので。
まあ、本場環境では 64bit 版にしてもらおう。*1
ちなみに、以下の条件だともっといい数字が出ている。（C クライアントを使った場合）
# USE_C_CLIENT=1 MESSAGES=400000 CONCURRENCY=40 DBI='dbi:mysql:test;mysql_socket=/tmp/mysql.sock;user=root;password=mysql' t/05-multireader.t
1..4
ok 1 - check number of messages
ok 2 - min value of received message
ok 3 - max value of received message
ok 4 - should have no rows in table
Multireader benchmark result:
Number of messages: 400000
Number of readers: 40
Elapsed: 23.972 seconds
Throughput: 16686.444 mess./sec.クライアントを C 版にするだけで３倍くらいのパフォーマンスになる模様。同じ条件で、perl 版クライアントを使った場合は、4970.574 mess./secだった。
