It needs only two load readings per request, yet in a simulation (see the post) it stays close to always picking the least loaded instance and far better than random.
Power of two choices
Load balancing rule: pick two random instances and send the request to the one with the shorter queue.