While running benchmark ab with our tool we found that g_timer_skiplist_mtx is acquired very frequently but the contention is very low. We experinted by replacing this pthread_mutex_lock with spinlock. We achieve 7% speedup running ./ab -k -c 350 -n 20000 localhost:8080/ in our 16 core linux machine.
There is a patch proposed in https://bz.apache.org/bugzilla/show_bug.cgi?id=57399#c5 which avoids this lock in the listener (unless necessary, i.e. it should be more contented in this case). Could you please pass your test(s) with it?