Comment 23 for bug 1011792

Hi Mike,

could you post the dmesg of that instance? Or actually if it is running for a while, boot messages may be gone from the ring buffer. Probably "sudo grep -r . /sys/hypervisor" in the guest is good enough.

So the issue was already there with Natty (2.6.38) but happens more often since Oneiric (3.0)/Precise (3.2). It sounds a bit problematic to create the same setup. At least I would not really know how to configure PostgreSQL and then create the necessary clients to hammer it.

Generally from the traces provided by Matt in comment #8, there seems to be quite a bit going on doing process management. Like balancing the cpu load. I have not gone through all the 16 cpu traces (it is a bit tedious) but a few of them seem to be on task structures. I will post info when I find something.

Meanwhile, even with the latest kernels not solving this, it would be interesting to find out whether it might be something (else) related to autogrouping. If you put the following either in /etc/sysctl.conf or create a new /etc/sysctl.d/60-autogroup.conf and add it there:

kernel.sched_autogroup_enabled = 0

Then reboot and check whether cat /proc/sys/kernel/sched_autogroup_enabled is indeed 0 before running the workload. That should show whether it has effects on the issue.