[Bug 2020] New: mysql multi-user performance half that of Fedora 2.4.22-2149.nptl kernel

[Martin J. Bligh]

http://bugme.osdl.org/show_bug.cgi?id=2020

           Summary: mysql multi-user performance half that of Fedora 2.4.22-
                    2149.nptl kernel
    Kernel Version: 2.6.2
            Status: NEW
          Severity: normal
             Owner: axboe@xxxxxxx
         Submitter: yusufg@xxxxxxxxxxxx


Distribution:

Fedora Core 1 (all updates applied)

Hardware Environment:

Dual P3-450 with 1GB RAM. 4x9GB SCSI disks attached to onboard Adaptec
aic7896/97 Ultra2 SCSI adapter (2 disks per channel) on a 440GX motherboard

Software Environment:

mysql 4.0.16 compiled from source
2.6.2 booted with elevator=deadline

Problem Description:

I received a benchmark program from Peter Zaitsev of Mysql. unlike the sql-bench
program in the mysql distribution which is single-threaded, this benchmark is
multi-threaded

I made some minor modifications to the source of sysbench.c (diff+original to be
attached) and ran it.

Comparing the number of transactions shown on 2.6.2 versus Fedora
2.4.22-2149.nptl kernel. 2.6.2 shows half the performance

Steps to reproduce:

compile insert.c and sysbench.c (to be attached)

The compile script for sysbench.c is

INC=/usr/local/site/mysql/include/mysql
LIB=/usr/local/site/mysql/lib/mysql

gcc -g -Wall -O6 -funroll-loops -fomit-frame-pointer sysbench.c  $LIB/libmysqlcl
ient.a  -lz -lm -lrt -o sysbench -I${INC} -L${LIB}

Replace INC,LIB as appropiate

Create a database 'bench'. Create table as per the following SQL command

CREATE TABLE `sbtest` (
  `id` int(10) unsigned NOT NULL auto_increment,
  `k` int(10) unsigned NOT NULL default '0',
  `c` char(120) NOT NULL default '',
  `pad` char(60) NOT NULL default '',
  PRIMARY KEY  (`id`),
  KEY `k` (`k`)
);

Create values to be inserted by running insert and saving the output in a file
./insert > insert.sql
insert the values into the mysql database

mysql -uroot bench < insert.sql

Run sysbench (my diff uses the mysql advanced test, the simple test is just a
trivial select which shows peak number of trivial queries you can get on that box)

./sysbench qt 

Observe the results (to be attached) 
vmstat 1 output during the tests is also attached


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/