Design and Implementation of a Compiler and Runtime System for Composite Tree Parallelism

Datatype (MAD) [Hud92], based on a functional construct called a Monad. This group has also developed a parallel programming model based on these constructs that allows explicitly parallel programs to be written that operate on these functional mutable data types [JH93]. This model uses a programming structure that is equivalent to our intermediate form to describe parallel computations in a referentially transparent, easily analyzable form. The drawback of this approach is that it requires the programmer to understand the underlying monadic formalism in order to write effective programs. The parallel mechanisms provided by this mechanism are also currently primitive message passing structures, and are difficult to assemble into high level parallel data structures. 3 The Composite Tree Model Data parallel programming languages are sequential languages that have been extended to access a parallel memory, in which parallel data structures can be stored. The conventional data parallel languages provide a parallel array data type, and a set of three basic operations that can be performed on the parallel array: selection, which produces a new parallel array from the subportion of the original array that meets some condition, mapping, which allows some primitive (usually arithmetic) operation to be performed on all elements of an array simultaneously, and reduction, which uses some arithmetic operation to combine the elements of an array into a scalar value. The basic model for our language was developed by taking the basic features of the data parallel model, and altering them in ways that make them better suited to general purpose programming on common coarse grained architectures. We make the following changes to the data parallel model: 1. Increase the granularity: Instead of using single arithmetic operations as the primitive grain of parallelism, we use entire object methods (member functions). 2. Make parallel data structures more flexible: Instead of restricting programmers to a single parallel array structure, we allow programmers to implement new parallel data structures in terms of a parallel tree abstraction. We do this by extending the standard single level host process/parallel memory model of most data parallel languages and architectures into a hierarchy. Tree structure is used, rather than a more general graph structure,