I know this is an old and probably unsupported version, but everyone is scared to touch these IBM boxes lest they explode.
I'm looking for some confirmation that this makes some kind of sense. I have multiple VMs, all running FreeBSD 9.3. All but two are very lightly loaded. All disks are local (2.5" 10K SAS). I noticed the other day when trying to do a repair on a mysql table that it was 1) slow and 2) was basically taking the other highly-loaded VM down. After some futzing around, I saw that about 50-70% of the CPU on the "slow" VM was consumed by the OSes interrupt process (seen in "top -S", and also in the CPU summary in "top"). Networking was affected as well - ping times would shoot up to 500-3000ms (from 0.2ms). If I suspended the VM that was doing the table repair, everything went back to normal. Resume it, the other host grinds to a halt again, drowning in interrupts. Both had very spotty disk IO performance in all cases - tab-completion in the shell was slow, even a simple "ls" was slow. The mysql VM was doing about 500 IOPS.
So I came up with two "fixes":
- Giving the mysql VM fewer disk shares did not change the behavior, but limiting the IOPS to 150 or so seemed to remedy things, but the mysql repair was even slower
- Not even sure how I came up with this, but I noticed the VM with the high interrupts had a CPU reservation of 4GHz - on a whim, I removed that and then removed the IOPS restriction on the mysql VM - all is well without that CPU reservation.
Does this make any sense? Am I just running into some weird edge case here?
