Towards a framework for cost-based transformation

Abstract

Program transformation systems are always based on some underlying model of cost. In sequential computation, this cost system is intuitive and hence almost transparent. In parallel computation this is not the case, primarily because of properties of communication, such as latency and congestion. These are hard to model abstractly because they depend on details of the decomposition of computations into threads, their placement on processors, and the way in which communication is handled in the interconnect. We present a framework for classifying cost models and the transformation systems they induce, and illustrate in the context of three parallel programming models. All restrict the flexibility of the programming system or limit the architectures to which they apply. The best trade-off between such restrictions and the expressiveness of the transformation system is an important open problem. The framework suggested here may be useful in the search for other useful cost models.