иконопис(BTW: still in London, then off to Stockholm, then home)
A customer just saw this with RHEL 6. Windows performance was higher than Linux performance on the same machine. The customer didn’t understand it, we guessed at first about it, and in the end our initial guess was wrong. But we caught what was wrong, with a WAG, and it troubled me. So I wrote this.
First clue as to the nature of the problem came from numastat. If you run numastat on a numa machine, you should get something like this:
landman@metal:~$ ssh jr5-lab numastat
numa_hit 11488947 17672711
numa_miss 0 0
numa_foreign 0 0
interleave_hit 27443 27648
local_node 11463950 17669861
other_node 24997 2850
Two nodes, about equal access between them. Mostly hitting the right nodes.
Now a machine with a single node either has one CPU populated with RAM (and the other without), or is configured not to run NUMA at all.