Modeling the distributed termination convention of CSP

Reviewer: Robert Elliot Filman

Communication Sequential Processes (CSP) [1], a language for distributed programming, uses guarded Boolean expressions to control intraprocess repetition and interprocess communication. In the original CSP, an attempt by a process, A, to communicate with a terminated process could then lead to the exiting of a loop in A. This mechanism of ending a loop in a process on the basis of the termination of another process came to be known as the distributed termination convention. Such a convention is a weakness in a distributed language, for it implies a level of global knowledge in a distributed system. The authors show that the distributed termination convention can be modeled in CSP by making processes, whose termination is relied on, send explicit terminat- ion messages to those processes that rely on their termination. Thus, if process X relies on the termination of process Y to exit a loop, Y should send X an explicit end signal before terminating. This paper shows that the addition of these explicit signals is a purely syntactic operation on the text of a CSP program. The distributed termination convention can thus be removed by an action akin to compiler optimization. (The solution relies on the use of output guards in CSP, a feature not in the original language definition, but used by most subsequent researchers.) I see two main weaknesses in the algorithm as described. The first is that it depends on the ability to syntactically determine communication partners in CSP. Even this analysis is subverted by the facility in CSP for indexing into an array of processes. Thus, the result does not generalize to languages that allow more flexible communication structures, such as the creation of new processes or the passing of process names in communications. Second, since the algorithm requires explicit code and communications for each potential reliance on termination, the size of this code and the number of communications can grow proportional to the square of the communicating processes. This will be particularly unfortunate for those programs that exploit process indices—syntactic analysis does not reveal the true communication structure of such systems.