srw_lock in log_sys.append_prepare() turned out to yield best throughput.
We might try a NUMA aware spin lock implementation later.