[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Inode/Blocksize questions



On Mon, 29 Apr 2002, Theodore Ts'o wrote:

> For files as small as single mail messages (i.e., averaging about
> 8-9k, rarely as larger than half a meg except for Windows virus
> messages), the performance gain for using a 4k blocksize is probably
> not all that great.  So you might want to use a 1k blocksize just to
> minimize the overhead.

Ok.

Here are some completely unreliable (only one run) bonnie++ benchmark
results:

1kb blocksize:

Version 1.02a       ------Sequential Create------ --------Random Create--------
mx                  -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete--
files:max:min        /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP
      50:10000:1000   241  96 35413  91 23396  90   239  99  8933  28   570  92


2kb blocksize:

Version 1.02a       ------Sequential Create------ --------Random Create--------
mx                  -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete--
files:max:min        /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP
      50:10000:1000   292  96 40624  90 25936  88   303  99 10096  28   817  92

So there's actually a noticable performance difference between 1kb and 2kb
block size.

> > With default options, the number of inodes mke2fs creates is 1 per 8kB of
> > disk space.
>
> In other words, mke2fs creates enough inodes for the case where the
> average size of a file on the filesystem is larger than 8k.  If you
> think you will be receiving a large number of small messages, it might
> be wise to set the bytes-per-inode setting (via the -i option of
> mke2fs) to 4096.  This will mean the filesystem will have enough
> inodes even if the average file size is as small as 4k.

Well, the default seems fine to me (7553013 inodes on a ~53 gig
partition). Our old mailserver which only has a 15 gig maildir storage
uses only ~500000 inodes.

> On the other hand, note that there have been some reports about
> reseirfs not handling I/O errors and disk corruption very well.
> Personally, I consider reliability far more important than
> performance, especially for a filesystem containing e-mail....

TRUE! We've got no battery-backed cache (*sighs*) on our ICP Vortex
controllers; reiserfs blew up into little pieces while doing some "what
happens if i pull the plug" tests (don't have the error messages anymore).
A readonly (!) reiserfsck resulted in a unusable partition while ext3
moved some stuff into lost+found and let all the other data on the
partition untouched. My personal experiences with reiserfs where
devestating.

best regards,
Michael Renner





[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]