Ubuntu ext4 file system hangs/slow at random times

[nate]

Hello folks -

I have an unusual problem on some Ubuntu VMs that seems to happen 
randomly.
I've seen it on maybe 3-4 different VMs at this point, on about 8 
occurrences
in the past couple of months. The only solution out of the issue is to 
power
cycle the VM.

One of the symptoms is very high system cpu usage with low I/O usage.

Software: Ubuntu 10.04.3 64-bit (2.6.32-37-server)
File system: ext4 w/LVM

Hypervisor: ESX 4.1 Update 2
Storage: the volume can either be a VMFS volume or a raw device mapped 
LUN
          from our fibre channel SAN (3PAR F200), all
          storage is thin provisioned.

Drivers: Using the paravirtualized SCSI adapters in ESX

The behavior is that at a random time the ext4 file system seems to get 
stuck.
Any process accessing the file system gets really slow access, and gets 
stuck
in a 'D' state. Underlying I/O performance is good with both service 
times
and average wait times under 1 millisecond. In one situation at least I 
tried
to do a tail on a sub 100 byte file during this sort of behavior and it 
hung
immediately.

Kernel dumps out tons of messages saying things were waiting longer 
than 120
seconds to run.

I have spent a least a couple hours searching on this topic but have 
not found
much information. I enabled ext4 event tracing in the
/sys/kernel/debug/tracing/set_event and those events are here:

http://yehat.aphroland.org/files/ubuntu/

I also put a snapshot of iostat running too.

I/O activity that has triggered this has varied:

- First time I saw it was when I was doing a parallel rsync of a few 
hundred gigs
   of data. I thought I worked around   it by disabling file system 
barriers.
   I noticed that at least in the RHEL 6.0 technical notes they require 
barriers.
   to be disabled when running enterprise storage (which we are). Our 
storage has
   mirrored cache and is battery backed.
   This volume was a raw device map.
- Basic log rotation of medium sized log files from a VMFS-based ext4 
volume to
   a NFS volume
- RRDtool activity from cacti to a raw device mapped volume
- Basic Splunk search/indexer activity(what I saw today).

3 different systems that I can think of off the top of my head. In 
every case the
VM in question has at least two different virtual disks, and only one 
of the
virtual disks is affected. The other one (root partition) is not.

Now that I know disabling barriers doesn't help I have moved two of the 
three
systems to be ext3 instead didn't re-format just remounted). I'm not 
sure whether
or not Ubuntu uses barriers by default on ext3 as well or just ext4.

If the solution is to stick to ext3 I have no problem with that - 
nothing in
ext4 that I need that I can think of really.

Though it would be nice if there was some fix to the issue.

I have about 150 VMs spread over 8 hosts connected to the same storage. 
All of
the VMs are managed the same way so they all have the same software and 
stuff.

If there is any other debug data I could gather that would be useful 
next
time this happens (assuming ext3 didn't fix it), please let me know.

This is my first experience with Ubuntu + ext4 + LVM on an enterprise 
storage
array. Not that I expected much of a different experience from 
CentOS/RHEL
(v4 and v5) + ext3 + LVM on the same array technology which I had been 
using
for years w/o issue.

thanks

nate