Sheep in wolf's Clothing: Implementation Models for Dataflow Multi-Threaded Software

Abstract

Concurrent programming is notoriously difficult, especially in constrained embedded contexts. Threads, in particular, are wildly nondeterministic as a model of computation, and difficult to analyze in the general case. Fortunately, it is often the case that multi-threaded, semaphore-synchronized embedded software implements high-level functional specifications written in a deterministic data-flow language such as Scade or (safe subsets of) Simulink. We claim that in this case the implementation process should build not just the multi-threaded C code, but (first and foremost) a richer model exposing the dataflow organization of the computations performed by the implementation. From this model, the C code is extracted through selective pretty-printing, while knowledge of the data-flow organization facilitates analysis. We propose a language for describing such implementation models that expose the data-flow behavior (the sheep) hiding under the form of a multi-threaded program (the wolf). The language allows the representation of efficient implementations featuring pipelined scheduling and optimized memory allocation and synchronization. We show applicability on a large-scale industrial avionics case study and on a commercial many-core.

Modeling and Control of bittide Synchronization

“…Synchronous programming models led to synchronous programming languages, e.g., Esterel [36], Lustre [37], Signal [38], and the development of tools to formally analyze their execution correctness as well as compilers to generate correct synchronizing code for embedded [2] or multicore platforms [4]. This created a virtuous cycle -as researchers understood better properties and embedded them into languages and tools, they drove the adoption of synchronous execution and formal tools for a number of industrial control applications, avionics, and critical system components.…”

“…Often, synchronous abstractions are decoupled from implementation and are used to validate system functional behavior. When mapping synchronous abstractions to asynchronous nondeterministic hardware, work has been done to automate code generation that matches the functional semantics, hiding the non-deterministic behavior of the hardware with explicit synchronization, for example [4]. Logical Execution Time (LET) was introduced by Henzinger and Kirsch [5] to support the design of reactive, cyberphysical systems.…”

Polyhedral Dataflow Programming: A Case Study

“…Although there is a significant amount of research for deriving efficient control for PPNs with particular constraints coming from the hardware [32], all this process remains polyhedral. Like the synchronous proposition of [33], we believe that these polyhedral works should be carefully integrated in our setting in order to gain benefit of both task and instruction parallelism; however we do not want to restrict the input language to a particular subclass of statically optimizable kernels, like in [34].…”