In the presence of significant latency or significant work to guard boundaries, throughput (read: appropriate architecture) matters, not performance (read: spending a week to shave off 50 CPU cycles).