[Samba] best filesystem choice for samba

[José Ildefonso Camargo Tolosa]

Hi!

Malcolm Baldridge wrote:

>Quoting Mark Lidstone <mlidstone at bmtseatech.co.uk>:
>
>  
>
>>ARGH!  I'm wondering if airing thoughts about VFAT performance publicly
>>was a good idea.
>>    
>>
>
>I doubt VFAT's case insensitivity would be worth dealing with its terrible
>linear-search-time directory lookup methods.
>
>The reason I suggested reiserfs (or ext3 with directory hashing) is to
>reduce the high costs of locating a directory entry within a directory of
>many (> 10,000) files.
>
>msdos/vfat does not offer superior directory lookup times, and from my
>limited testing, neither does NTFS.
>
>ext2/ext3 in stock configuration is also slow, though it appears very recent
>kernels/ext2fsutils offer an FFS-like "directory hashing" option which needs
>a format-time decision to be made upon setting up the filesystem.
>  
>
You can enable it with tune2fs:

obelix:~# tune2fs -O dir_index /dev/hda3

See man tune2fs for more help.

> <>
> I have no knowledge about XFS or JFS and how they compare. I know both are
> "industrial" filesystems brought down from the Ivory Towers onto the
> pipsqueak platforms.
>
> As for "horror stories", well, each filesystem has had their respective
> tales of misery and woe... ext3 had shocking and fatal dataloss bugs 
> in the
> adolescent versions of 2.4.x., and some RAID + reiserfs configs saw some
> real wowsers as well. From bug reports/changelogs, I've seen similar tales
> of woe for XFS and JFS if you trigger just the right combination of 
> things.
> >From my own experiences, things have matured and stabilised with reiserfs
> and ext3 to the point where using either is fine for my purposes.

I had very bad experience with reiser: 4 servers installed with reiser, 
4 server died due to filesystem corruption in a time that varied from 
two to six months  (the last one had UPS, the others not).  I 
reinstalled them with ext3: almost a year since I reinstalled the first: 
no problems.

>The decision comes down to:
>
>1) Do you need quotas?  If yes, you cannot use reiserfs.
>2) Do you need ACLs?  If yes, only ext2/ext3 has well-tested seamless
>support, though I think there are wildcat patches to bring this to XFS (and
>maybe others) as well.  I'm not sure about the stability of this.
>
>ext3 used with -O dir_index *MAY* provide better performance for large
>directory list lookups, but I've never tested it.  It requires Linux 2.6 for
>starters for the kernel-side stuff to actually support it properly. 
>grepping the linux 2.4 source shows no mention of hashing b-trees or
>dir_index options for ext[23].
>
>This is a RECENT addition to ext3, and I don't think the support actually
>exists within 2.4 yet.  I've seen mention of "special backported patches"
>but this smells scarier to me than using filesystems which have been
>seamlessly integrated for over a year or so now.
>
>So in terms of viable performance-driven alternatives, I see it being
>reiserfs, xfs, or jfs.
>  
>
In my experience: the fourth server (the one with the ups): Dual XEON 
2Gb RAM, 3x36Gb scsi disk in raid-5 array smart array 5300, running 
squid: it was slower then (with reiser), than now (with ext3).  I have 
only saw reiser to be faster when I delete a LARGE file (>1Gb).  I'm 
going to test ext3 with the dir_index option.

>vfat/dos isn't faster, even with case insensitive semantics, for directory
>sizes of 20,000 or more.
>  
>
I agree.

Ildefonso
icamargo at unet.edu.ve
icamargo at merkurio.com.ve
ildefonso_camargo at yahoo.com