Graphical interfaces and windowing are now the norm for computers-human interaction. Also, advances in computer networking have given computer users access to immense distributed resources accessible from anywhere on the network. In this setting, the desktop, or personal computer plays the role of a user-interface engine that mediates access to the available resources. Interface paradigms, such as the "desktop metaphor" and "direct manipulation," provide the user with a consistent, intuitive view of the resources. Traditional computer research has focused on enhancing computer performance from the numerical processing and transaction processing perspectives. In the research described in this thesis a systematic framework is developed for analyzing and improving the performance of window systems and graphical user interfaces. At the system level a protocol-level profiling strategy has been developed to profile the performance of display-server computers. A sample protocol-level profiler, Xprof, has been developed for applications under the X Window System. At the micro- architecture level the memory access characteristics of windowing programs are studied. Cache tradeoffs for a frame-buffer cache are presented. A cache organization is proposed to improve the frame-buffer performance.