My system is a Debian 6.0.3 x86_64 box w/kernel 2.6.32-5-openvz-amd64, it hosts a number of OpenVZ containers. I have recently migrated from LXC/newer kernel from backports repo to the stable OpenVZ kernel to resolve network issues but got another issue – the CPU spends a lot of time processing software interrupts. Here is a Munin graph for CPU time: http://prntscr.com/arjzl.
I added nohz=off and highres=off to the kernel command line but that did not help much. And, the number of interrupts is quite low on this box according to another Munin graph: http://prntscr.com/ark19, so the interrupts seem to be not the cause of this issue. When I run top, zabbix_server is the most CPU consumer but I don't see any relation between zabbix_server and software interrupts.
I'm not sure how to proceed with investigation, this looks like a kernel bug or a kernel module bug to me, but I don't know hot to track this down to the guilty process/module. Maybe there is another kernel command line parameter to tune. What should I try next?
Best Answer
I wonder how exactly do you try seeing a relation? Have you tried stopping it, for example?
Since this problem has arisen after kernel change, it looks reasonably to assume this is a bug of OVZ 2.6.32-5. I suggest using http://wiki.openvz.org/Download/kernel/rhel6/042stab055.16 if it's fresher than yours current (as it seems).