Ext3 users - Dec 2003 - Ext3 on RAID5 for video archive server

I am using a terabyte RAID5 with ext3 and created using mdadm, to store jpeg
files on a rehdat 7.3 system. I need to write files to the RAID5 at the rate of
800 files per second, and each file is about 10-20 KB in size.
 
  I have tried using data-writeback and the noatime options while mounting, and
I am also using a stipe-size of 4MB (since 4MB was the max that mdadm would
allow me), and i have passed -R stride=1024, and -b 4096 to mkfs.ext3 while
formatting the RAID5.
  Still, I keep getting "hiccups" every few seconds while writing the
files.. mostly, these hiccups are at 5-second intervals, so it may suggest that
the ext3 default commit interval is the culprit, and the effect has diminished
after i changed the commit interval to 1 second in the kernel code, under
linux/fs/jbd/journal.c, the line changed was:
journal->j_commit_interval = (HZ * 1); // originally HZ * 5.

    I have some questions at this point:
1. Is there any other way of modifying the commit interval? If possible I would
like to take the commit interval modification out of the kernel and put it back
in user level code..
2. Are there other things that I should try to optimize the write performance?
Thanks a lot in advance!!!


---------------------------------
Do you Yahoo!?
New Yahoo! Photos - easier uploading and sharing

Hi,

On Wed, 2003-12-17 at 15:06, M K wrote:
>   Still, I keep getting "hiccups" every few seconds while writing
the
> files.. mostly, these hiccups are at 5-second intervals, so it may
> suggest that the ext3 default commit interval is the culprit, and the
> effect has diminished after i changed the commit interval to 1 second
> in the kernel code, under 
> linux/fs/jbd/journal.c, the line changed was:
> journal->j_commit_interval = (HZ * 1); // originally HZ * 5.
>     I have some questions at this point:
> 1. Is there any other way of modifying the commit interval? If
> possible I would like to take the commit interval modification out of
> the kernel and put it back in user level code.. 
Yes, you can mount with "-o commit=1" (or put "commit=1" in
the options
field in /etc/fstab) as long as you've got a moderately recent kernel. 
The 2.4.20-20.* Red Hat Linux kernels should include that.
> 2. Are there other things that I should try to optimize the write
> performance?
There are various things you could try.  In general, avoiding
dependencies between different IOs is a good thing.  That can mean
splitting up files amongst more directories (htree will help to avoid
the cost of large dirs, of course); or even splitting the storage into
multiple different filesystems, or splitting the raid array into
multiple arrays.  Really, though, it's so application-dependent that
it's hard to be any more specific than this.

Cheers,
 Stephen