Thursday, November 14, 2013

Using ParaSail as a Modeling Language for Distributed Systems

The ACM HILT 2013 conference just completed in Pittsburgh, and we had some great tutorials, keynotes, and sessions on model-based engineering, as well as on formal methods applied to both modeling and programming languages.  One of the biggest challenges identified was integrating complex systems with components defined in various modeling or domain-specific languages, along with an overall architecture, which might be specified in a language like AADL or SysML or might just be sketches on a whiteboard somewhere.  A big part of the challenge is that different languages have different underlying semantic models, with different type systems, different notions of time, different concurrency and synchronization models (if any),  etc.  The programming language designer in me wants to somehow bring these various threads (so to speak) together in a well-defined semantic framework, ideally founded on a common underlying language.

One way to start is by asking how can you "grow" a programming language into a modeling language (without killing it ;-)?  ParaSail has some nice features that might fit well at the modeling level, in that its pointer-free, implicitly parallel control and data semantics are already at a relatively high level, and don't depend on a single-address-space view, nor a central-processor-oriented view.  As an aside, Sebastian Burckhardt from Microsoft Research gave a nice talk on "cloud sessions" at the recent SPLASH/OOPSLA conference in Indianapolis (,, and we chatted afterward about what a perfect fit the ParaSail pointer-free type model was to the Cloud Sessions indefinitely-persistent data model. Modeling often abstracts away some of the details of the distribution and persistence of processing and data, so the friendliness of the ParaSail model to Cloud Sessions might also bode well for its friendliness to modeling of other kinds of long-lived distributed systems.

ParaSail's basic model is quite simple, involving parameterized modules, with separate definition of interface and implementation, types as instantiations of modules, objects as instances of types, and operations defined as part of defining modules, operating on objects.  Polymorphism is possible in that an object may be explicitly identified as having a polymorphic type (denoted by T+ rather than simply T) and then the object carries a run-time type identifier, and the object can hold a value of any type that implements the interface defined by the type T, including T itself (if T is not an abstract type), as well as types that provide in their interface all the same operations defined in T's interface.

So how does this model relate to a modeling language like Simulink or a Statemate?  Is a Simulink "block" a module, a type, an object, or an operation (or something entirely different)?  What about a box on a state-machine chart?  For Simulink, one straightforward answer is that a Simulink block is a ParaSail object.  The associated type of the block object defines a set of operations or parameter values that determine how it is displayed, how it is simulated, how it is code-generated, how it is imported/exported using some XML-like representation, etc.  A Simulink graph would be an object as well, being an instance of a directed graph type, with a polymorphic type, say "Simulink_Block+," being the type of the elements in the graph (e.g. DGraph).

Clearly it would be useful to define new block types using the Simulink-like modeling language itself, rather than having to "drop down" to the underlying programming language.  One could imagine a predefined block type "User_Defined_Block" used to represent such blocks, where the various display/simulation/code-generation/import/export operations would be defined in a sub-language that is itself graphical, but relies on some additional (built-in) block types specifically designed for defining such lower-level operations.  Performing code-generation on these graphs defining the various primitive operations of this new user-defined block type would ideally create code in the underlying programming language (e.g. ParaSail) that mimics closely what a (ParaSail) programmer might have written to define a new block type directly in the (ParaSail) programming language.  This begins to become somewhat of a "meta" programming language, which always makes my head spin a little...

A practical issue at the programming language level when you go this direction is that, what was a simple interface/implementation module model, may want to support "sub-modules" in various dimensions.  In particular, there may be sets of operations associated with a given type devoted to relatively distinct problems, such as display vs. code generation, and it might be useful to allow both the interface, and certainly the implementation of a block-type-defining module to be broken up into sub-modules.  The ParaSail design includes this notion, which we have called "interface extensions" (which is a bit ambiguous, so the term "interface add-on" might be clearer).  These were described in:
but have not as of yet been implemented.  Clearly interface add-ons, for say, [#display] or [#code_gen], could help separate out the parts of the definition of a given block type.

A second dimension for creating sub-modules would be alternative implementations of the same interface, with automatic selection of the particular implementation based on properties of the parameters specified when the module is instantiated.  In particular, each implementation might have its own instantiation "preconditions" which indicate what must be true about the actual module parameters provided before a given implementation is chosen.  In addition, there needs to be some sort of a preference rule to use when more than one implementations' preconditions are satisfied by a given instantiation.  For example, presume we have one implementation of an Integer interface that handles 32-bit ranges of integers, a second that handles 64-bit ranges, and one that handles infinite range.  Clearly the 32-bit implementation would have a precondition that the range required be within +/- 2^31, the 64-bit one would require the range be within +/- 2^63, and the infinite-range-handling implementation would have no precondition.  If we were to instantiate this Integer module with a 25-bit range, the preconditions of all three of the implementations would be satisfied, but there would presumably be a preference to use the 32-bit implementation over the other two.  The approach we have considered for this is to allow a numeric "preference" level to be specified when providing an implementation of a module along with the implementation "preconditions," with the default level being "0" and the default precondition being "#true." The compiler would choose the implementation with the maximum preference level with satisfied preconditions.  It would complain if there were a tie, requiring the user to specify explicitly which implementation of the module is to be chosen at the point of instantiation.

No comments:

Post a Comment