Re: Removing softdep

[Ignatios Souvatzis <ignatios%cs.uni-bonn.de@localhost>]

Hi,

On Wed, Jun 11, 2008 at 10:38:20AM +0300, M.K. wrote:

> I think having a journalling filesystem is a good thing, as far as it remains
> compatible with older installations. But why not use LFS? It has been much
> improved recently and should do the job, although  it surely has its
> drawbacks. Since I'm not a filesystem expert, my question is: What is the
> basic difference between using ffs+journalling and using LFS? I'm using
> LFS and so far I haven't experienced any noticeable problems. I don't
> say LFS doesn't have any bugs, most probably it does, but its performance is
> good enough. Can someone give more information about how it behaves in server
> environment under heavy load?

LFS uses the (free segments (== a fixed small number of blocks) of) the
partition as a big circular buffer of memory for write operations,
instead of the usual FFS style allocation.

Think of it as a fully (metadata and data) journaled filesystem, but
there is only the journal.

As free blocks stay blocked when other parts of a segment are still in
use, a cleaner process runs around the segments, and moves the used
parts of partially used segments to the end of the log, freeing their
original location.

I was a big fan of LFS on slight load 4.0-ish machines but moved away
because they started to sort of lock up under certain conditions.

What happens is that the cleaner can't free more segments than it needs,
so it never finishes one cycle, and all writers wait for a clean segment.

I was advised to increase kern.maxvnodes as a vnode is how the kernel
holds onto yet-unwritten changed data and forces writing of a segment
even if it isn't full if it runs out of vnodes, but couldn't find a
value that allowed for reliable - I don't remember: system or pkgsrc -
builds on my systems.

Regards,
        -is