Tuning Linux for Databases?

Malcontent asks: "Since it's possible to tweak Linux to a high degree I wonder what tricks can be employed to make Linux a better database server. I am not talking about obvious hardware improvements like RAID, more processors or memory because they apply to any operating system. I am looking more for Linux specific things like kernel settings and disk/filesystem tricks which may improve performance."

Filesystem Tricks...

[lw54]

I am looking more for Linux specific things like kernel settings and disk/filesystem tricks.

The biggest filesystem trick is to not use a filesystem.

Get a real database and access the disks directly. :-)

DB Tuning

[pbryant]

The following are some lessons I've learned the painful way. They may be of some use. 0) Do you really want to do this? Hire an expert, consult with your vendor. 1) Get the latest version of Linux. We ran into such horrible problems with Red Hat 6.0 and Oracle I still have nightmares. 6.2 works a lot better. Redhat are now selling an 'oracle8i' tuned Linux version for $2100. 2) Increase your file limits. echo "16384" >/proc/sys/fs/file-max echo "49152" >/proc/sys/fs/inode-max ulimit -Hn 16384 (as root) ulimit -n 16384 (as your db user) From here, its standard db tuning: 3) Memory is a biggie. Make sure you have a ton. With Oracle I use the following rule of thumb: count the number of DIMM slots in your machine, multiply that by the largest DIMM size available e.g. 512Mb. You should be able to scrape by on this. 4) Disks: the more the better. Use straight disks (no raid), Raid 0, raid 1, or raid 0+1 (depending on the level of safety you need), we use 0+1. Raid 5 has a write performance impact. The Oracle redo logs are a bottleneck. We write out two copies of our log to 2 different straight disks for max performance. 5) More CPU won't hurt.