Section: New Results

Static Analysis of OpenStream Programs

Participants : Albert Cohen [Inria Parkas team] , Alain Darte, Paul Feautrier.

In the context of the ManycoreLabs project, we started to study the applicability of polyhedral techniques to the parallel language OpenStream [19]. When applicable, polyhedral techniques are indeed invaluable for compile-time debugging and for generating efficient code well suited to a target architecture. OpenStream is a two-level language in which a control program directs the initialization of parallel task instances that communicate through streams, with possibly multiple writers and readers. It has a fairly complex semantics in its most general setting, but we restricted ourselves to the case where the control program is sequential, which is representative of the majority of the OpenStream applications.

In contrast to the language X10, which we studied in previous years, this restriction offers deterministic concurrency by construction, but deadlocks are still possible. We showed that, if the control program is polyhedral, one may statically compute, for each task instance, the read and write indices to each of its streams, and thus reason statically about the dependences among task instances (the only scheduling constraints in this polyhedral subset). If the control program has nested loops, communications use one-dimensional channels in a form of linearization, and these indices may be polynomials of arbitrary degree, thus requiring to extend to polynomials the standard polyhedral techniques for dependence analysis, scheduling, and deadlock detection. Modern SMT allow to solve polynomial problems, albeit with no guarantee of success; the approach previously developed by P. Feautrier [14], and recalled in Section 7.1, offers an alternative solution.

The usual way of disproving deadlocks is by exhibiting a schedule for the program operations, a well-known problem for polyhedral programs where dependences can be described by affine constraints. In the case of OpenStream, we established two important results related to deadlocks: 1) a characterization of deadlocks in terms of dependence paths, which implies that streams can be safely bounded as soon as a schedule exists with such sizes, 2) the proof that deadlock detection is undecidable, even for polyhedral OpenStream. Details of this work have been published at the international workshop IMPACT'16 [1].

Some further developments are in progress for scheduling OpenStream programs using polynomial techniques (with a corresponding prototype scheduling tool, specific to OpenStream, see Section 6.3). In particular, we made some progress for parsing a simplified version of OpenStream, exhibiting the relevant structure, and on the properties and construction of schedules with bounded streams and bounded delays, and on the analysis of the “foot bath”, i.e., the pool of tasks that are created (already requiring some resources) but not activated yet (because they need to wait for the termination of other tasks due to dataflow semantics). This work should have interesting connections with the way runtime systems of tasks are managed.