Subject: Re: Why my life is sucking. Part 2.
To: None <current-users@netbsd.org>
From: Alan Barrett <apb@cequrux.com>
List: current-users
Date: 01/17/2001 18:26:38
On Wed, 17 Jan 2001, Herb Peyerl wrote:
> not using softdep.  About to not use raid.

For a fair test of whether raidframe is really at fault, or whether
it's actually FFS having trouble with certain combinations of size and
geometry, I suggest that you try copying relevant parameters from the
raid disklabel to a non-raid disklabel.

I haven't had any raid trouble, except when a hardware timeout
eventually caused a crash.  Recovery worked fine, but after "raidctl
-R" had reconstructed the data onto the mirror, "raidctl -P" was also
necessary.  Surely, in the case of a simple two-way RAID1 mirror, both
"raidctl -R" and "raidctl -P" simply copy data from the good disk to the
other disk, and there should be no need to copy the same data twice?

I did a few slightly non-standard things with my raid1 setup:

  * Put several raid sets on the same pair of disks, ensuring that each
    raid set was smaller than 8GB.  This was out of sheer paranoia in
    case large raid sets didn't work properly.

  * Adjusted the raid disklabel's sectors/track, tracks/cylinder,
    sectors/cylinder and cylinders fields because the defaults
    looked weird.  I made the disklabel sectors/track value equal to
    the raidframe sectPerSU value (128 in my case), and increased
    tracks/cylinder to 32 (from the default of 1).

  * Ensured that each filesystem was fairly small (no larger than 2GB).

  * Told newfs to use a larger than default number of cylinders per
    cylinder group.

As a consequence of the above, I have filesystems with relatively few
cylinder groups, where each cylinder group is relatively large.  This
would make it much less likely for me to run into the ffs_dirpref()
problem when a cylinder group gets full.

--apb (Alan Barrett)