- Trade off granularity to reduce overhead, since the current model
doesn't appear to reduce contention substantially: move to a single
harvest mutex protecting harvesting queues, rather than one mutex
per source plus a mutex for the free list.
- Reduce mutex operations in a harvesting event to 2 from 4, and
maintain lockless read to avoid mutex operations if the queue is
full.
- When reaping harvested entries from the queue, move all entries from
the queue at once, and when done with them, insert them all into a
thread-local queue for processing; then insert them all into the
empty fifo at once. This reduces O(4n) mutex operations to O(2)
mutex operations per wakeup.
In the future, we may want to look at re-introducing granularity,
although perhaps at the granularity of the source rather than the
source class; both the new and old strategies would cause contention
between different instances of the same source (i.e., multiple
network interfaces).
Reviewed by: markm
full, avoiding the cost of mutex operations if it is. We re-test
once the mutex is acquired to make sure it's still true before doing
the -modify-write part of the read-modify-write. Note that due to
the maximum fifo depth being pretty deep, this is unlikely to improve
harvesting performance yet.
Approved by: markm
can more easily be used INSTEAD OF the hard-working Yarrow.
The only hardware source used at this point is the one inside
the VIA C3 Nehemiah (Stepping 3 and above) CPU. More sources will
be added in due course. Contributions welcome!
