[Lustre-discuss] lfsck

[Andreas Dilger]

On Jun 04, 2008  14:37 -0400, Charles Taylor wrote:
> Here is a question (or perhaps a veiled comment/complaint).     Why  
> does lfsck list *every* file name to stderr?    Doesn't that just make  
> a painfully slow process even slower?    Would it be better to just  
> report progress periodically and any *important* errors that need to  
> be addressed.
> 
> So far on a 33TB file system (8 OSSs, 24 OSTs), the time required has  
> been....
> 
> Generate mdsdb:    6.5 hours  (94GB, sparse)
> Generate ostdbs:    45 mins each (400 to 600 MB each)
> Copying db's   around :   45 min to an hour
> Actual LFSCK:   ?????  (an hour so far but we have 30,600,000 files  
> (roughly) so I fear this will get ugly.
> 
> I'm not making any real point here - just sharing information and,  
> well, killing time... :)

We've been trying to get rid of lfsck for some time now, having this
checking done internally to Lustre.  Unfortunately, as the battle of
features goes, this one usually loses out.  The net result is that
we don't want to spend too much time on a "dead" tool in preference
to making the new implementation, so lfsck gets very little attention.

Once we go to ZFS, lfsck will be dead for sure and the new (internal
verification) mechanism will have to be implemented.

Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.