The interaction network : a performance measurement and evaluation tool for loosely-coupled distributed systems

摘要

Much of today's computing is done on loosely-coupled distributed systems. Performance issues for such systems usually involve interactive performance, that is, system responsiveness as perceived by the user. The goal of the work described in this thesis has been to develop and implement tools and techniques for the measurement and evaluation of interactive performance in loosely-coupled distributed systems. The author has developed the concept of the interaction network, an acyclic directed graph designed to represent the processing performed by a distributed system in response to a user input. The definition of an interaction network is based on a general model of a loosely-coupled distributed system and a general model of user interactions. The author shows that his distributed system model is a valid abstraction for a wide range of present-day systems. Performance monitors for traditional time-sharing systems reported performance information, such as overall resource utilisations and queue lengths, for the system as a whole. Performance problems are now much more difficult, because systems are much more complex. Recent monitors designed specifically for distributed systems have tended to present performance information for execution of a distributed program, for example the time spent in each of a program's procedures. In the work described in this thesis, performance information is reported for one or more user interactions, where a user interaction is defined to be a single user input and all of the processing performed by the system on receiving that input. A user interaction is seen as quite different from a program execution; a user interaction includes the partial or total execution of one or more programs, and a program execution performs work as part of one or more user interactions. Several methods are then developed to show how performance information can be obtained from analysis of interaction networks. One valuable type of performance information is a decomposition of response time into times spent in each of some set of states, where each state might be defined in terms of the hardware and software resources used. Other performance information can be found from displays of interaction networks. The critical path through an interaction network is then defined as showing the set of activities such that at least one must be reduced in length if the response time of the interaction is to be reduced; the critical path is used in both response time decompositions and in displays of interaction networks. It was thought essential to demonstrate that interaction networks could be recorded for a working operating system. INMON, a prototype monitor based on the interaction network concept, has been constructed to operate in the SunOS environment. INMON consists of data collection and data analysis components. The data collection component, for example, involved the adding of 53 probes to the SunOS operating system kernel. To record interaction networks, a high-resolution global timebase is needed. A clock synchronisation program has been written to provide INMON with such a timebase. It is suggested that the method incorporates a number of improvements over other clock synchronisation methods. Several experiments have been performed to show that INMON can produce very detailed performance information for both individual user interactions and groups of user interactions, with user input being made through either character-based or graphical interfaces. The main conclusion reached in this thesis is that representing the processing component of a user interaction in an interaction network is a very valuable way of approaching the problem of measuring interactive performance in a loosely-coupled distributed system. An interaction network contains a very detailed record of the execution of an interaction and, from this record, a great deal of performance (and other) information can be derived. Construction of INMON has demonstrated that interaction networks can be identified, recorded, and analysed.