mkfs.xfs with a 9TB realtime volume hangs

[Dave Chinner]

On Sat, Jan 10, 2009 at 03:13:23PM -0600, Eric Sandeen wrote:
> Jan Wagner wrote:
> > Hi,
> > 
> > I have a RAID0 with 11x750GB+1x1TB components in the following 
> > partitionable-md test setup
> > 
> > root at abidal:~# cat /proc/partitions | grep md
> >   254     0 9035047936 md_d0
> >   254     1     124983 md_d0p1
> >   254     2    1828125 md_d0p2
> >   254     3    1953125 md_d0p3
> >   254     4 9031141669 md_d0p4
> > 
> > Essentially, four partitions: 128MB, ~1.9GB, 2GB, 9TB. I'd like to use the 
> > 1.9GB partition for xfs and put a realtime subvolume onto the same raid0 
> > onto the 9TB partition. The partition tables are GDT instead of MBR to be 
> > able to have >=2TB partitions.
> 
> Sorry for the slow/no reply.  It seems to be doing many calculations in
> rtinit, haven't sorted out what yet, but it's not likely hung, it's
> workin hard.  :)
> 
> If you give it a larger extsize it should go faster (if the larger
> extsize is acceptable for your use...)
> 
> I tried a 4t realtime volume:
> 
> mkfs.xfs -dfile,name=fsfile,size=1g -rfile,name=rtfile,size=4t,extsize=$SIZE
> 
> for a few different extent sizes, and got
> 
> extsize	  time
> -------	  ----
> 512k	  0.3s
> 256k	  0.7s
> 128k	  1.9s
>  64k	  8.4s
>  32k	 25.4s
>  16k	129.4s
> 
> With the default 4k extent size this takes forever (the man page claims
> default is 64k, maybe this got broken at some point).

It got changed a few years back by Nathan, IIRC. I bet the time
being taken a result of the blow-out in bitmap size caused by reducing
the extent size. Given it is non-linear, it may have something to do
with cache sizes as well.  e.g buftarg hashes not large enough.

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com