[Lustre-discuss] WRT "obdidx ordering in "lfs getstript"
Andreas Dilger
adilger at whamcloud.com
Thu Feb 9 21:31:22 PST 2012
On 2012-02-09, at 6:49 PM, Ms. Megan Larko wrote:
> I was reading Mr. David's query about the ordering of data on a
> striped luster file system. I too am under the impression that the
> data stripe of size lfs-stripesize will rotate in order from the
> starting point. Following Mr. David's example, a large data set
> would be written to the 2nd OST, with the next piece on the 3rd, then
> 0th and finally 1st before circling back around to the 2nd (assuming
> OSTs 0 to 3 from the example).
Correct.
> In his response, Mr. Dilger stated:
> "when OST free space is imbalanced the OSTs will be selected in part
> based on how full they are". Does that refer to a starting point for
> the data writes before the orderly progression? Does that somehow
> imply a "skipping over" of a "full" OST?
Correct. When free space becomes too imbalanced between OSTs, the MDS
object allocator changes to a mode where it allocates objects partly
based on how much space is free on each OST. This is not ideal, and
could be improved (see https://bugzilla.lustre.org/show_bug.cgi?id=18547
for details), but is reasonable for some workloads.
> The latter would be
> revolutionary to me in my personal understanding of Lustre and cluster
> file systems in general. I thought that a single OST having
> insufficient space available for writing of the data piece of "stripe
> size"---or all of the data if the default Lustre stripe size of one is
> used--would cause a file system full error. This error can confuse
> users and novice administrators who see a file system full message
> when a typical disk usage command on the client will show (ofter a
> reasonable) percentage available on the file system as a whole.
This is still true after a file has had OSTs allocated. If any OST
becomes full, writes to files on that OST will return ENOSPC even if
there is free space on another OST.
Cheers, Andreas
--
Andreas Dilger Whamcloud, Inc.
Principal Engineer http://www.whamcloud.com/
More information about the lustre-discuss
mailing list