On Mon, May 2, 2016 at 11:54 AM, Jan Engelhardt
On Monday 2016-05-02 16:41, Greg Freemyer wrote:
The scheduler bugs seem to affect systems using the NUMA feature of the scheduler more than those that don't.
lscpu | grep NUMA
provides information about NUMA. Are the servers in the public instance of OBS using NUMA based hardware. i.e. Do they have more than one NUMA node?
Theoretical approach: A worker pool is something that is commonly scaled horizontally rather than vertically, usually because of cost per work and redundancy feature set. Finding a system with more than one node is therefore a bit unlikely, and one with more than two is virtually nonexistent in such a setting.
But, many of the physical machines are running 12 or more VMs. Not likely small machines. A few Intel based servers (build21, build24, and build27) currently are configured to run 16 VMs (per the status monitor display). PPC machines build67, build91 and build92 also have 16 VMs configured. I would not at all be surprised if those 6 machines had a NUMA architecture.
Practical approach: Run a build job with lscpu in it, and find out. (Though, virtualization may mask the real value and always yield 1.)
I'm assuming the virtualization would indeed mask the real value. Someone with SSH access to the physical machines would need to run the command "lscpu | grep NUMA". Greg -- To unsubscribe, e-mail: opensuse-buildservice+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-buildservice+owner@opensuse.org