On Tue, 31 Mar 2015, Chris Metcalf wrote:
> This may come back to a question of just why one believes that
> nohz_full is a good thing in the first place. For folks that are doing
> it just to improve performance, power, etc, generally, it may not
> matter much whether the watchdog ticks occasionally. But for folks
> who are doing it to establish cores that are run completely tick-free
> for days on end so they can help process 100 Gb packet streams
> and never drop a packet, the calculus is a little different. My bias
> is to say that once you've tagged a core as nohz_full, you never want
> to run the watchdog on it. But it's worth supporting multiple uses
> of nohz_full, certainly.
Completely tick-free is some ideal. What is really needed is to make it
rare. We do not want to cripple the cores that have a reduced tick.
If someone configures a watchdog to run every X seconds on a NOHZ core
then he should get the tick every X seconds because that is what he
requested. If someone else do not want the watchdog to run then they are
willing to accept the limitations in debugging that come with it.
For that to work we need to have a way to configure this. Maybe allow the
specification of a cpumask somewhere?