On Sun, Sep 09, 2012 at 08:23:20PM +0530, Uma Maheswara Rao G wrote:
Hi Experts,
I am using Linux for my production clusters. I know When we access the disk concurrently, performance will go bad. (ext3 file system)
What are the best practices to avoid this problem? one point I read to tune the schedulers. Other than this, If there are any best practices, Please share.
I am trying from my application to reduce the concurrent access.
Hallo Uma, I am pretty sure you're OT on this ML, despite the fact, you use hadoop (we even do not distribute). All tuning you have to perform is kernel, so please contact the kernel community http://en.opensuse.org/Portal:Kernel Anyway - you should be way more specific about your problem as there might be different strategies. Basically you can * use tune2fs to change filesystem settings * use deadline and tune it's parameters * buy more expensive and powerfull HW (ie with SSD) * ... The best answer heavilly depends on the typical workflow of an application. For instance disabling barriers won't help you with reads. There you can benefit from having some often read things on SSD dics. so if you do not expect an answer from /dev/crystall-ball (/dev/oracle is probably copyrighted ;-)), you'd need to provide them more informations. What is the typical workload of your application - concurrent read, writes, read&writes. Are those operations spread along many files or partitions, or not. What's the settings of ext3 you have or tryed? Which io scheduler do you use (deadline, cfq) and which seetings is there. What is the HW you use for storage. On which kernel version do you run. Regards Michal Vyskocil