Turns out a large amount of thread contention was coming from these shared counters -- disable some of them and add some thread-local caching