[lopsa-discuss] Re: Linux 2.6 vs 2.4 [was Reserve memory in Linux RHEL3 & 4]

[Theodore Ts'o]

On Wed, Feb 15, 2006 at 11:29:56AM -0500, Marc Staveley wrote:
> The biggest problem I have with Linux 2.6 is that I can't use it for
> write intensive applications.  kjournald just eats the machine for a
> significant period of time (10's of seconds) after a big write.

How big of a write are we talking about?  And this is something that
doesn't happen with 2.4?  That seems surprising, since there haven't
been any major changes in the journalling algorithm between 2.4 and
2.6.

Something that may or may not help in this case, but in general seems
to make ext3 run more smoothly is to user a larger journal that the
default chosen by mke2fs.  Recent version of e2fsprogs now use a
substantially bigger journal; the default used to be 32MB, I now
recommend a journal size of 128MB for most filesystems.

> I also have noted a major difference in the software RAID behaviour.
> Under 2.4 a mirror resync of a 120GB disk takes about 90 minutes, and
> the machine is very usable during the resync.  Under 2.6 (same machine
> and disks) the resync takes 26 HOURS, with kjournald eating 80% of the
> CPU and the machine is totally unusable until the resync finishes.

Reconstruction shouldn't be affecting kjournald, except that of course
a lot of your disk throughput is going to the MD
reconstruction/resync.  What is the application that you're running on
your server?  It sounds like it is extremely metadata intensive....

						- Ted