[lustre-discuss] stripe count recommendation, and proposal for auto-stripe tool

[Mohr Jr, Richard Frank (Rick Mohr)]

> On May 19, 2016, at 12:46 PM, Nathan Dauchy - NOAA Affiliate <nathan.dauchy at noaa.gov> wrote:
> 
> Thanks for pointing out the approach of trying to keep a single file from using too much space on an OST.  It looks like the Log2(size_in_GB) method I proposed works well up to a point, but breaks down in the capacity balancing department at some large file size.  If we take your rule of thumb that no file should use more than 1% of an OST, and a typical new-ish target made up of an 8+2 RAID10 group of 6TB disks, we should divide by 480 GB rather than 150.  (are your targets a couple years old? or am I making a bad assumption here?)

Yes, my targets are several years old.  The 1% number I used is just something I thought was reasonable.  It is arbitrary, and you can choose whatever makes the most sense for your site.

> Since the size of OSTs is a moving target (pun intended), it would be good to "future proof" the size selection method.  An automated tool like "lfs_migrate -A" could actually calculate the changeover point from log2() your method on the fly.  stripe_count=MAX( Log2(size_in_GB)+1, size_in_GB/X ) where "X" is ~1% of the capacity of the smallest OST.  This strays away from simplicity for the user, but the tool hides it for restriping, and initial writes are (hopefully) tuned more for concurrent access pattern than size anyway.

That’s true.  If your restriping will mostly be done by a script, then you don’t necessarily need a simple formula.

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