On Tue, 2012-12-11 at 09:02 -0500, Steven Rostedt wrote:

Currently, what we have is a huge contention on both the pulled CPU rq
lock. We've measured over 500us latencies due to it. This hurts even the
CPU that has the overloaded task, as the contention is on its lock.

The 500us latency was one of the max ones I saw, and I believe it was
combined with the load balancer going off at the same time as the pulls
were happening. I've seen larger latencies with limited function tracing
enabled and the load balancer was very involved.

As I believe we have users that would very much want this in, and if you
are still not sure you like this "feature" I can make it into a
SCHED_FEAT() like TTWU_QUEUE. Then you can keep this off and the big
boxes can enable them.

Maybe enable the feature by default if the box has more than 16 cpus?

-- Steve