Structural Clocks for a Loosely Synchronized Data-Parallel Language

Nowadays, most of distributed architectures are MIMD (Multiple Instruction streams, Multiple Data streams) par - allel computers or networks of workstations. They are eas- ily scalable, concerning the number of processing elements , the amount of memory or the bandwidth of the underlying communication network. The challenge consists in taking advantage of the power of these architectures. It has been shown that data-parallel languages offer both a program- ming model easy to understand, and several execution mod- els which are able to exploit these distributed architectur es. The compilation process has to fill the gap between the syn- chronous and centralized programmer's point of view and an asynchronous and distributed execution model. An im- portant problem is to reduce the number and the complexity of synchronizations. Our contribution consists in the def- inition of an intermediate asynchronous data-parallel lan - guage where synchronism becomes explicit in the syntax. Its semantics is based on a twin memory management. We pro- pose an execution model where synchronizations are par- tial and handled through the use of structural clocks. These clocks record the position of each processing element with regard to the program structure. They are a generalization of Awerbuch's synchronizers. The relevance of the approach is illustrated by examples.