Saturday, July 08, 2006

More info on iSCSI benchmark

Looks like a lot of interest was generated from the results of RightMedia's iSCSI benchmarking. Peter Zaitsev is curious about the environment of the benchmark. So here's all the gory details
MySQL version : 5.0.21
System : GenToo Linux, kernel version 2.6.16-gentoo-r2-1
Hardware : 4GB memory, 2 Dual Core AMD Opteron 275 chips
MySQL configuration file (abridged version):

[client]
port = 3306
socket = /var/run/mysqld/mysqld.sock

[mysqld]
basedir = /usr
datadir = /var/lib/mysql/data
port = 3306
pid-file = /var/run/mysqld/mysqld.pid
log-error = /var/log/mysql/mysqld.err
socket = /var/run/mysqld/mysqld.sock
user = mysql
max_connections = 1024
table_cache = 128
thread_case_size = 32
sort_buffer_size = 64M
innodb_additional_mem_pool_size = 20M
innodb_buffer_pool_size = 3000M
innodb_data_home_dir = /var/lib/mysql/inno
innodb_log_group_home_dir = /var/lib/mysql/log/inno
innodb_log_file_size = 2000M
innodb_log_files_in_group = 2
innodb_log_buffer_size = 8M
innodb_flush_trx_at_commit = 2
innodb_autoincrement_size = 32M
innodb_status_file
innodb_thread_concurrency = 4
innodb_flush_method = O_DIRECT
innodb_support_xa = 0

sysbench v0.4.3
Command used to prepare 100,000,000 record sbtest database:
sysbench --num-threads=1 --test=oltp --mysql-user=root --oltp-table-size=100000000 --mysql-table-type=innodb prepare

For each SysBench run:
sysbench --num-threads=# --test=oltp --mysql-user=root --oltp-table-size=100000000 --mysql-table-type=innodb --max-requests=10000 --oltp-test-mode=complex run

Before each SysBench run, the kernel pagecache was also cleaned out:
sync ; echo 1 > /proc/sys/vm/drop_caches

This was done to provide a fair & clean test. For each batch of runs by thread, 4 runs were made,with the best TPS used as the final number for the batch.

I am pretty happy with the results. I plan on putting for iSCSI benchmark results soon, as I think a lot of folks would be interested in the results, especially if the iSCSI device scales well.

3 comments:

Anonymous said...

Thank you for clarifying results.
I still would like to see some VMSTAT printout however.

Looking at parameters you used you used right database but you used default access distribution which is very skewed - in majority of cases rows from very small primary key range will be accessed this is why performance is that high and difference is that insignificant.

It does not mean iSCSI will do bad for you in real life, just make sure you tested with load similar to what you will have in real life.

Speaking about clearing caches - this is one of the ways you can do it but this makes benchmark results unstable in the sense - 1min benchmark will return different results from 15min benchmark. I usually prefer "hot" benchmark with warmup done and only "steady state" measured. This is pretty much as TPC-C or SpecJAppServer run as well.

Happy tuning. If you publish any more benchmarks let me know. I'm always hungry for new results.

Partha Dutta said...

Unfortunately I had to give up my testing boxes to put into production use. On the iSCSI systen, a "prepare" run of 10,000,000 records took approximately 8 minutes. This is with a 3GB InnoDB buffer pool, and innodb_flush_method=O_DIRECT

Anonymous said...

So to have a good attack, you should level your skill in cabal online alz. Also, get the highest level jewelry that you can equip in thecabal alz. This step depends on your skill rank, for lower magic ranks get 4 fast notice arrow skills and your cabal gold. Of course, if you have more cabal money, you can get more in the game. By the way, you can buy pet from other players in buy cabal alz.
Many people play the game maybe feel that they were too poor, some of my friend often ask, why you play games so there is ghost gen.
Say good pull so much I do not pull hedge with the US to talk about the game experience to make ghost online gen. I want to see if there is nothing worth ghost gold things and a very low price. The most important thing is you have ghost online gold. Also we have a good time happy in the game and hope earned more ghost money in the game.