In the context of Differentiated Services (DiffServ), we investigate the effect of acknowledgment marking on the throughput of TCP connections. We carry out experiments on a testbed offering three classes of service (Premium, Assured and Best-Effort), and different levels of congestion on the data and acknowledgment path. We apply a full factorial statistical design and deduce that marking TCP data packets is not sufficient and that acknowledgment marking on the reverse path is a necessary condition to efficiently match targeted performance in DiffServ. We find that the optimal marking strategy depends on the level of congestion on the reverse path. In the practical case where Internet Service Providers cannot obtain such information in order to mark acknowledgment packets, we show that the strategy leading to optimal overall performance is to copy the mark of the respective data packet, provided that the affected service class is appropriately provisioned.