Low-latency versus throughput