A general but simple technique to handle asynchronous data-parallel control structures

Nowadays, most of distributed architectures are eas- ily scalable MIMD (Multiple Instruction streams, Multiple Data streams) parallel computers or networks of worksta- tions. The challenge consists in taking advantage of the power of these architectures. It has been shown that data- parallel languages offer both a programming model easy to understand and several execution models which are able to exploit these distributed architectures. The compilation pro- cess has nevertheless to fill the gap between the synchronous and centralized programmer's point of view and an asyn- chronous and distributed execution model. An important problem is to reduce the number and the complexity of syn- chronizations between processors. This leads to complex program transformations, such as absorption techniques, i n conditioning control structures, loops or routines. We pro - pose a simple general technique based on structural clocks that permits the use of one-sided wait instructions instead of global synchronizations. Hence, no program transfor- mation is needed and synchronization requirements are re- duced. In particular, management of recursive routines is greatly facilitated. To illustrate these techniques, we pr o- pose an application to Strassen's algorithm for matrix prod - uct.