K 10
svn:author
V 3
mav
K 8
svn:date
V 27
2012-12-14T21:22:23.463798Z
K 7
svn:log
V 503
Add some simple adaptiveness to the maximal next even lookup depth.

If callout_process() was called two times within the same millisecond,
limit lookup by 1/16s instead of default 1/2s.  On the amd64 test system
it reduces usleep(1) time from 8us to 6us, handling about 180K callouts
and thread wakeups per second in one stream with not full CPU load.

That is the only price of staying with the callwheel data structure.
Unless at some point we want to sleep/lookup for minutes, it doesn't
look high.

END