This thesis explores the problem of making accurate assessments of the performance of high level language interpreter programs which are embedded in some more complex system. The overall system performance will be determined by all the software and hardware components present; but in order either to analyse and improve particular components, or to select between alternative versions of components, the concept of the performance of individual components is important. A model is developed for the abstract behaviour of software components playing the role of an interpreter by considering their interaction with the program code which is being interpreted and with the underlying virtual machine which is, in turn, interpreting them. This model enables a flexible definition of performance by relating the interactions in which an interpreter takes part. A methodology is recommended for assessing experimentally the performances defined within such a framework. The performances of an interesting selection of pseudo-machine and high level interpreter implementations of Lispkit and Prolog are then assessed and conclusions drawn.