Currently only one thread is serving the selector. Under high load it is possible that the selector thread could get overloaded and become the bottleneck. It might be interesting to explore if the burst scheduling overhead drops when there are more threads working with the selector.