On 02/05/11 09:32, Jean-Francois wrote:
> Hello,
> 
> I just read some extracts of a paper, study from Margo Seltzer & Keith A. 
> Smith from Harvard university, a comparison of LFS & FFS.

the paper from 1995??

Dude.  That's a LONG time ago in the computer world.  It is also a very
non-specific "Log-structured file system", which may or may not have any
real-world counterpart here 16 years later (yes, some modern file
systems are "logging" FSs, but...are they descendants of this 1995 LFS?
 Or was this LFS a dead-end for real-world reasons that never show up in
academic papers?  (I'm sure I could do some more research on this, but
it's your question, not mine :)

> Basic questions from my side, is FFS-2 better than FFS in the sense of 
> dealing 
> with creation of many small files, and is fragmentation less than with FFS ?

Please describe the fragmentation problem you have /observed/...  I do a
lot to torment file systems, and never seen anything that looked like a
PROBLEM caused by fragmentation on OpenBSD.  If you aren't seeing a real
problem, how can you benefit from optimizing?

> Are other file systems with some improvement of performance compared to FFS 
> available for OpenBSD ?

Short answer: there are two file systems provided for day-to-day use on
OpenBSD: FFS and FFS2.  FFS is the general purpose OS, FFS2 is for very
large file systems which can't be handled by FFS.  Nice and simple.

Other file systems that OpenBSD supports are for cross-system
compatibility, not for "better" anything on OpenBSD, at least at this
time (wouldn't mind seeing a working HAMMER port, of course).

And...as FFS2 is used for larger file systems, I think it is safe to say
that putting lots of small files on huge file systems is much worse than
putting lots of small files on a few (or a lot) of small file systems.

However, if you are looking at writing lots of small files, make sure
you you are using softdeps, you will get a very large performance gain
(I'm not talking 10% -- more like 10x!).  You may find you get much
better real performance than many logging systems give.

Nick.

Reply via email to