filesystem performance problem

[West, Brian]

I have a problem in a 400GB ext3 filesystem that is mapped to a Raid 10
lun on an IBM DS4800. The filesystem was created using the defaults of
mkfs. I'm using RedHat AS 4.
 
The problem is that performance tails off to about 2 MB / sec. when
writing to this filesystem. It starts out fast for several seconds,
upwards of 100MB /sec. and then tails off. This condition can be
reproduced in many ways:
1. oracle experiences same problem
2. copying large database files from another filesystem into the problem
filesystem
3. using dd to write a 5 GB file. Sometimes it'll work and the entire
file will get created but most of the time it slows down to a crawl. 
 
Nothing else is running on the server or on the storage system.
 
What's more bizarre is that the problem only seems to be in a directory
under the mount point of the problem filesystem. If I write directly to
the mount point it's fine. And, if I go a few directories beyond the
problem directory it is fine.
 
We seem to have gotten around the problem by renaming the directory,
creating another directory and copying the files back into the new
directory.
 
Other filesystems created the same way on other servers using the same
storage system are fine.
 
Any ideas ?