Re: The maximum number of files under a folder

["Stephen Samuel" <darkonc gmail com>]

The OS will have to search the directory to see if the file already exists before creating it.

Well, if you hash it such that it splits up something like:
jobid(upper part)/jobid(lower- part)[/-]timestamp-process,

 you'll find that your access times will be must faster (especially if
you don't use H-Trees).  This also applies if  you're just creating a
file, because you'll have to search the entire directory to see if that
filename exists

With regular directories, searching through them to see if a file
already exist increases linearly with the number of entries.  If you
hash on 3 levels with 8-bits per level, you'll have to open 2 or 3
extra inodes, but you'll cut your directory search times down by a
factor of 20000-1.  You'll also skip having to deal with any sort of
directory-size limit. (=2^24/256/3)

I did something similar on a Solaris box which had 200000 emails in
the /var/spool/mqueue directory. That many messages was slowing the
system to a crawl.  I hashed it into 100 directories with 2000  entries
each,   it sped things up enormously.

On Tue, Mar 18, 2008 at 3:56 PM, Andreas Dilger <adilger sun com> wrote:

On Mar 17, 2008  09:32 -0400, Theodore Ts'o wrote:
> On Mon, Mar 17, 2008 at 03:40:36PM +0800, liuyue wrote:
> > Theodore Tso,
> >
> >     In 64bit system, directory size can not be bigger than 2GB?
>
> No, because the high 32-bits for i_size are overloaded to store the
> directory creation acl.

I think we should change the code (kernel and e2fsprogs) to allow
i_size_high for directories also.

> In practice, you really don't want to have a directory that huge
> anyway.  Iterating through it all with readdir() gets horribly slow,
> and applications that try do anything with really huge directories
> would be well advised to use a database, because they will get *much*
> better performance that way....

Actually, for many HPC applications they never do readdir at all.
The job creates 1 file/process and always uses a predefined filename
like {job}-{timestamp}-{process} that it will directly look up.

Cheers, Andreas

-- 
Stephen Samuel http://www.bcgreen.com
778-861-7641