“PEWMA is opportunistic in that it tries to get the best latency, and this means it may sometimes leave a server less than fully loaded.”
Round robin works when servers are equal and requests cost the same, which is never. Queues cut dropped requests and pay for it in latency. PEWMA shifts latency lower than least connections, then drifts worse over time, and ships with a pile of tunable parameters. The best algorithm here wins by leaving capacity idle, and someone pays for that idle box.