Return network traffic (from servers to clients) in thin-client systems is modeled as a mixture of interactive data flows corresponding to keystrokes and bulk data flows related to screen updates. Users are very sensitive to delay and jitter of the former flows. Thus our goal is to minimize the latency of interactive data transfer without increasing latency of bulk data transfer. Through simulation experiments, we determine that the main factors causing end-to-end delay in the interactive data transfer are queuing delay in the router and buffering delay in the server. When we apply two techniques: priority queuing of interactive data flows at the router and using TCP SACK option, the average end-to-end delay can be reduced. However, several servers could take more than a second to send large bulk data flows; this delays the transmission of following interactive data flows. We then develop TCP optimization mechanisms: modifying recalculation of the retransmission timeout value and temporarily turning off the TCP SACK control, and demonstrate that they can overcome the negative effects of the existing techniques.