Segmented Key Cache Performance
Contents
Testing method for segmented key cache performance
We used SysBench v0.5 from Launchpad to test the segmented key cache performance for the MyISAM storage engine of MariaDB 5.2.2-gamma.
As wrapper scripts for automated running of SysBench we used the sysbench/ directory from MariaDB Tools.
To test that splitting the key cache's global mutex into several mutex helps under multi user load, we wrote a new SysBench test called select_random_points.lua. We used one big table and selected random points with increasing number of concurrent users.
Main testing outcomes
We see up to 250% performance gain depending on the amount of concurrent users.
Detailed testing outcomes
On our machine pitbull
On pitbull with --random-points=10
In relative numbers:
Threads 1 4 8 16 32 64 128 (32/off) -3% 53% 122% 155% 226% 269% 237% (64/off) -6% 55% 130% 162% 234% 270% 253% select_random_points.lua --random-points=10
On pitbull with --random-points=50
In relative numbers:
Threads 1 4 8 16 32 64 128 (32/off) -3% 53% 113% 154% 232% 254% 231% (64/off) -1% 55% 121% 161% 235% 268% 244% select_random_points.lua --random-points=50
On pitbull with --random-points=100
In relative numbers:
Threads 1 4 8 16 32 64 128 (32/off) -3% 54% 121% 160% 209% 246% 219% (64/off) -6% 56% 129% 167% 219% 260% 241% select_random_points.lua --random-points=100
Detailed numbers of all runs on pitbull
You can find the absolute and relative numbers in our OpenOffice.org spread sheet here: SysBench v0.5 select_random_points on pitbull
On our machine perro
On perro with --random-points=10
In relative numbers:
Threads 1 4 8 16 32 64 128 (32/off) 1% 2% 17% 45% 73% 70% 71% (64/off) -0.3% 6% 19% 46% 72% 74% 80% select_random_points.lua --random-points=10
On perro with --random-points=50
In relative numbers:
Threads 1 4 8 16 32 64 128 (32/off) 1% 10% 26% 69% 105% 122% 114% (64/off) -1% 8% 27% 75% 111% 120% 131% select_random_points.lua --random-points=50
On perro with --random-points=100
In relative numbers:
Threads 1 4 8 16 32 64 128 (32/off) -0.2% 1% 22% 73% 114% 114% 126% (64/off) -0.1% 4% 22% 75% 112% 125% 135% select_random_points.lua --random-points=100
Detailed numbers of all runs on perro
You can find the absolute and relative numbers in our OpenOffice.org spread sheet here: SysBench v0.5 select_random_points on perro
Table and query used
Table definition:
CREATE TABLE sbtest ( id int unsigned NOT NULL AUTO_INCREMENT, k int unsigned NOT NULL DEFAULT '0', c char(120) NOT NULL DEFAULT '', pad char(60) NOT NULL DEFAULT '', PRIMARY KEY (id), KEY k (k) ) ENGINE=MyISAM
Query used:
SELECT id, k, c, pad FROM sbtest WHERE k IN (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
The ? parameters were replaced by random numbers when running the SysBench test. We used 10, 50, and 100 random points in our tests.
We inserted 20 million rows using random data, which gave us a data and index file size of:
3.6G sbtest.MYD 313M sbtest.MYI
We chose our key buffer size to be big enough to hold the index file.
Testing environment
MariaDB sources
We used MariaDB 5.2.2-gamma with following revision from our launchpad repository Revision #2878
revno: 2878 committer: Sergei Golubchik <sergii@pisem.net> branch nick: 5.2 timestamp: Tue 2010-10-26 07:37:44 +0200 message: fixes for windows
Compiling MariaDB
We compiled MariaDB using this line:
BUILD/compile-amd64-max
MariaDB runtime options
We used following configuration for running MariaDB
MYSQLD_OPTIONS="--no-defaults \ --datadir=$DATA_DIR \ --language=./sql/share/english \ --log-error \ --key_buffer_size=512M \ --max_connections=256 \ --query_cache_size=0 \ --query_cache_type=0 \ --skip-grant-tables \ --socket=$MY_SOCKET \ --table_open_cache=512 \ --thread_cache=512 \ --key_cache_segments=0 \ # 0 | 32 | 64 --tmpdir=$TEMP_DIR"
SysBench v0.5 select_random_points.lua options
We run the SysBench v0.5 select_random_points.lua test with following options:
# 20 million rows. TABLE_SIZE=20000000 SYSBENCH_OPTIONS="--oltp-table-size=$TABLE_SIZE \ --max-requests=0 \ --mysql-table-engine=MyISAM \ --mysql-user=root \ --mysql-engine-trx=no \ --myisam-max-rows=50000000 \ --rand-seed=303"
We tested with increasing number of concurrent users with a warm up time of 8 minutes and a run time of 20 minutes:
NUM_THREADS="1 4 8 16 32 64 128" ... --num-threads=$THREADS
We also tested an increasing number of random points:
# Default option is --random-points=10. SYSBENCH_TESTS[0]="select_random_points.lua" SYSBENCH_TESTS[1]="select_random_points.lua --random-points=50" SYSBENCH_TESTS[2]="select_random_points.lua --random-points=100"
Kernel parameters
IO scheduler
For optimal IO performance running a database we are using the noop scheduler. You can check your scheduler setting with:
cat /sys/block/${DEVICE}/queue/scheduler
For instance, it should look like this output:
cat /sys/block/sda/queue/scheduler [noop] deadline cfq
You can find detailed notes about Linux schedulers here: Linux schedulers in TPCC like benchmark.
Open file limits
Having a lot of concurrent connections can hit the open file limit on your system. On most Linux systems the open file limit is at 1024, which can be not enough. Please set your open file limit higher by editing
$EDITOR /etc/security/limits.conf
and adding a line like
#ftp hard nproc 0 #@student - maxlogins 4 * - nofile 16384 # End of file
Your ulimit -a output should look like this afterwards:
ulimit -a core file size (blocks, -c) 0 data seg size (kbytes, -d) unlimited scheduling priority (-e) 0 file size (blocks, -f) unlimited pending signals (-i) 15975 max locked memory (kbytes, -l) 64 max memory size (kbytes, -m) 1744200 open files (-n) 16384
Machines used for testing
perro
# OS: openSUSE 11.1 (x86_64) # Platform: x86_64 # CPU: Quad-core Intel @ 3.20GHz: 4 CPUs # RAM: 2GB # Disk(s): 2 x ST31000528AS S-ATA as software RAID 0
pitbull
# OS: Ubuntu 10.10 # Platform: x86_64 # CPU: Two-socket x hexa-core Intel Xeon X5660 @ 2.80GHz. With hyperthreading: 24CPUs # RAM: 28GB # Disk(s): 1 x ST3500320NS S-ATA
