For your busy-polling test, I'd also record quote age when the application makes its decision, not just request-response latency. A fast response can still carry a price that waited in a queue. Replay a burst at a fixed input rate and compare the oldest quote used, dropped messages and CPU before and after the change.