Re: [rfc] Near-constant time directory index for Ext2

[Daniel Phillips <phillips@innominate.de>]

"H. Peter Anvin" wrote:
> 
> Andreas Dilger wrote:
> >
> > Basically (IMHO) we will not really get any noticable benefit with 1 level
> > index blocks for a 1k filesystem - my estimates at least are that the break
> > even point is about 5k files.  We _should_ be OK with 780k files in a single
> > directory for a while.
> >
> 
> I've had a news server with 2000000 files in one directory.  Such a
> filesystem is likely to use small blocks, too, because each file is
> generally small.
> 
> This is an important connection: filesystems which have lots and lots of
> small files will have large directories and small block sizes.

I mentioned this earlier but it's worth repeating: the desire to use a
small block size is purely an artifact of the fact that ext2 has no
handling for tail block fragmentation.  That's a temporary situation -
once we've dealt with it your 2,000,000 file directory will be happier
with 4K filesystem blocks.  There will be a lot fewer metadata index
blocks in your directory file, for one thing.  Another practical matter
is that 4K filesystem blocks map directly to 4K PAGE_SIZE and are as a
result friendlier to the page cache and memory manager.

--
Daniel