[gpfsug-discuss] default inode size

Stephen Ulmer ulmer at ulmer.org
Wed Mar 15 13:36:24 GMT 2017



> On Mar 15, 2017, at 8:37 AM, Lukas Hejtmanek <xhejtman at ics.muni.cz> wrote:
> 
> On Wed, Mar 15, 2017 at 08:22:22AM -0400, Stephen Ulmer wrote:
>> You need 4K nodes to store encryption keys. You can also put other useful things in there, like extended attributes and (possibly) the entire file.
>> 
>> Are you worried about wasting space?
> 
> well, I have 1PB of capacity for data disks and 3TB of capacity for metadata
> (SSD).
> 
> With 512B inodes, this ration seemed to be quite ok, while with 4k inodes,
> I run out of free space on SSD pretty fast. So I'm thinking about re-creating
> file system with smaller inode size.
> 
> I don't think I will ever need encryption keys.


What’s the average file size? Every file that is less than *about* 3.5K will wind up in the (4k) inode itself.

Also check the HW sector size of the SSDs (many are 4K now). If that’s the case, you’ll actually be much more efficient from an access point of view with 4K inodes.

If you don’ t have 1PB of data yet (or won’t in the next year or so), you can always expand the system pool with more SSDs in the future.

It’s not scary to pick another size — it works just fine, but you’re about to create a filesystem that will be a pain in the rear to change. It will take a long time to move/backfill even a few hundred TB of data if you change you mind later. You’ll definitely give up (possibly future) features by choosing not 4K, but you’ve got the flexibility to do so. :)

Liberty,

-- 
Stephen




More information about the gpfsug-discuss mailing list