[NTLUG:Discuss] Filesystem Performance

[cbbrowne@ntlug.org]

Kenneth Loafman wrote:
> Looking for some expert help here.  We need to evaluate Linux vs Windows
> performance re filesystems, namely filesystems with a large number of
> files, say in the order of several million, half in the 3-5k range, half
> in the 40-50k range.  Do you know of any studies done that would give me
> some rationale (other than cost) to use Linux over M$.
> 
> At what point does the filesystem performance start to degrade with the
> shear number of files?  Smaller directories, smaller partitions, and
> tuning on both systems will help, but is there an upper limit on either
> one?
> 
> As usual, a pointer to a 'real' study would be of help.  This decision
> could mean thousands of dollars going to the Evil Empire if I cannot
> prove my point completely.

This is going to be _really_ thorny to work with; performance can vary
very substantially based on such things as:

-> What are the patterns of access?  Mostly reads?  Mostly writes?
-> What filesystems are you using?  NTFS?  Ext2?  Ext3?  ReiserFS?

I would be utterly unsurprised to hear a benchmark changing by a factor
of 2 based on fiddling with block sizes and such.

There's not a single fixed form of "performance;" metrics will include:
-> Speed of access to directory entries as the number of files in a single
   directory grows;

-> Speed of updates to file data;

-> Speed of updates to metadata (e.g. - directory info);

Speed will vary based on things such as:
 -> How full are the inode tables?
 -> How many files in one directory?
 -> How full is the disk?

Generally speaking, if the disk gets really full, Performance Will Suck.

And it's not entirely reasonable to evaluate things based on the assumption
that the disk's completely full; that establishes only that they're after
pathological performance characteristics.  (Note that DOS FAT gets Really
Bad when things get pathological...)

A perhaps _bigger_ deal, with a 24GB filesystem, is the issue of what
happens if the server gets powered down unexpectedly.  With ext2, it
would be not remarkably unsurprising to see this result in a 2 hour
fsck marathon.  You'll _really_ want to start looking at filesystems
with better resyncing characteristics, like ReiserFS.
-- 
Christopher Browne
<http://www.cbbrowne.com/info/>
cbbrowne at cbbrowne.com