The evacuation task queue needs to be improved:
- add proper padding to the queues themselves to avoid false sharing and improve repeatability of results
- the overflow task queue needs an overhaul, or at least its sizing improved: measurements with solaris studio analyzer showed a large amount of time spent in trying to allocate new or free used segments in the overflow task queue (in malloc/free); the current sizing indicates around 512 entries per segment, which is way too small for any reasonably large application.
Overhaul means to implement some kind of exponential segment size growth to decrease the overhead.
Fix this.