On Mon, Mar 21, 2005 at 07:40:53AM -0500, Joshua Baker-LePain wrote:
>On Sun, 20 Mar 2005 at 4:25pm, Robin Humble wrote
>> I hammered a standard fc3 kernel (4k stacks) and couldn't break it.
>> So that's one extra data point for a relatively simple config that says
>> XFS and 4k stacks is pretty stable.
>Hrm. I had exactly the opposite experience. My testbed was far older
>(and simpler) -- dual PIII 450, 384MB RAM, AIC-7890 controller and 2 SCSI
>disks (not in any sort of RAID or anything). I was running RHEL4 on it
>with the kernel modified simply to turn on XFS support.
>'tiobench --size 2047' would reliably produce stack overflows.
<insert another couple of days of testing>
I can't reliably produce stack overflows. I once crashed the machine
(no logs produced, could ping it but nothing else) with a 4k stack AS4
kernel with XFS when using few tiobench's in combination with a memory
thrashing app that grabbed all the ram it could. But other runs
survived just fine... So clearly there's a problem there, but it's rare
enough to be hard to reproduce with my hardware config.
The stock fc3 kernel (4k stack) with AS4 userland with tiobench + memory
thrashing seemed a little better than the AS4 kernel, and didn't have
any problems except one warning stack trace that (understandably)
complained about lack of memory.
Various other kernels which had 8k stacks (AS4 8k, recent XFS cvs)
worked fine with tiobench. Multiple tiobench's and over NFS too.
Perhaps your results can be explained if the aic7xxx driver is a lot
more stack-heavy than the 3ware driver that we are using?
Or if you are exercising multiple XFS partitions at the same time?