Subject: Re: [vserver] HASH FS
From: Herbert Poetzl <herbert@13thfloor.at>
Date: Fri, 8 Aug 2008 03:06:12 +0200

On Mon, Jul 21, 2008 at 11:49:51PM +0900, Jun OKAJIMA wrote:
> ( This mail is sent to both OpenVZ and vserver ML)
> 
> Hello folks.
> 
> This is a new FS for virtual server.
> Basically, Ths is "vhashify/vzcache" FS.
> Just try it, please. I need your feed back.
> 
> URL:
> http://www.digitalinfra.co.jp/20080720/hashfs.20080720.html

okay, took me some time to answer, but here it is:

first, I do not consider FUSE to be a good choice
for production systems and/or anything which should
be performant and reliable, but, and I guess that is
the case here, it is a good tool to do prototyping

second, as I stated in the email you reference below,
I do not consider unification something you want to
do at the filesystem layer, more something you want
on the vfs layer ... and that is why we implemented
the CoW Link Breaking some time ago ...

third, what worries me most in your Overview is that
you ignored the fact that nowadays unification is
not done to save disk space, but to reduce memory
usage (by reducing caches and mappings), which is
something the Hash FS doesn't do and probably won't
be able to do for the foreseeable future ...
(the key to this would be reusing the inode numbers
wherever you have shared data)

finally, I think you should consider 'automatic'
hashing and re-combination of files, but that probably
makes the whole thing too complicated to maintained

on a completely different note, I think that hash
based filesystems are something very interesting,
especially if you create the hashes on a per block
basis and thus allow to use them for disk space
reduction as well as for data integrity checks

best,
Herbert

> Best regards,
> 
>          --- Okajima, Jun. Tokyo, Japan.
> 
> To Mr. Herbert Poetzl :
> 
> See this URL:
> http://www.mail-archive.com/vserver@list.linux-vserver.org/msg04247.html
> >> BTW, I also am planning a new file system for Vserver.
> >well, let's hear about it then ...
> 
> This is it.