On Mon, Jul 24, 2017 at 02:55:00PM -0600, Chris Murphy wrote: > On Mon, Jul 24, 2017 at 2:42 PM, Hugo Mills <h...@carfax.org.uk> wrote: > > > > > In my experience, it's pretty consistent at about a minute per 1 > > GiB for data on rotational drives on RAID-1. For metadata, it can go > > up to several hours (or more) per 256 MiB chunk, depending on what > > kind of metadata it is. With extents shared between lots of files, it > > slows down. In my case, with a few hundred snapshots of the same > > thing, my system was taking 4h per chunk for the chunks full of the > > extent tree. > > Egads. > > Maybe Cloud Admin ought to consider using a filter to just balance the > data chunks across the three devices, and just leave the metadata on > the original two disks? > > Maybe > > sudo btrfs balance start -dusage=100 <mp>
It's certainly a plausible approach, yes. Or just wait it out -- the number of slow chunks is typically very small. Note that most of the metadata will be csums (which are fast), and not all of the other metadata chunks are slow ones. It would be interesting to know in this case the times of the chunks that have been balanced to date (grep for the lines with the chunk IDs in system logs). Hugo. -- Hugo Mills | Two things came out of Berkeley in the 1960s: LSD hugo@... carfax.org.uk | and Unix. This is not a coincidence. http://carfax.org.uk/ | PGP: E2AB1DE4 |
signature.asc
Description: Digital signature
