Re: Reserved space (WAS: How to calculate bsdlabel size)

On  9 Feb, Bruce Evans wrote:

> The table also shows that disks much more than 50% full may be quite
> slow.  All real disks that I've tested are even slower than the table
> indicates (mainly due to large seeks to other cylinder groups, and
> large backwards seeks I believe).  E.g., my FreeBSD src tree can be
> written and read at almost half the maximum disk speed (20MB/s out of
> 40MB/s) when copied to a new partition, but reading it from its usual
> place goes at 3-5% of the maximum disk speed (1 or 2 MB/s) despite
> me copying my usr partition back and forth to defragment it less than
> a year ago.

The dirpref code was really badly broken and made poor decisions about
which cylinder groups should be used when new directories were created.
I fixed the worst of the brain damage about three months ago in rev
1.116 of ffs_alloc.c, but I think the algorithm could be further
improved.  In its current state, the code tends to overfill groups of
cylinder groups, while others have lots of free space.

It would also be helpful in these days of large disks and large files if
cylinder groups could be made larger, but they are limited by the
requirement that the cylinder group data structure size is the same as
the file system block size.  A possible workaround would be to cluster
groups of N cylinder groups when making allocation decisions.