[Lustre-discuss] Performance dropoff for a nearly full Lustre file system

[Mohr Jr, Richard Frank (Rick Mohr)]

On Jan 15, 2015, at 10:26 AM, Mike Selway <mselway at cray.com>
 wrote:

> 	Thank you for the explanation.  So, by definition ldiskfs will in a sense "hide" 5% of the disk capacity?  That is, if the usable disk space is 500TB, the system will hold aside  25TB?  If so, will the file system report 475TB usable and report 100% full at 475TB consumed?

I believe the "hidden 5%" is just the default reserved block percentage that ext file systems use when they are created.  If you run "df' on an ost, you should see that the "Used" plus "Available" blocks add up to 95% of the "1K-blocks" value.  This value can be changed using "tune2fs -m", but it is subject to debate on whether or not this is a good idea.  I decreased this to 1% on one of our Lustre file systems with no noticeable ill effect.

-- 
Rick Mohr
Senior HPC System Administrator
National Institute for Computational Sciences
http://www.nics.tennessee.edu