Musket: A Domain-Specific Language for High-Level Parallel Programming with Algorithmic Skeletons

Parallel programming for an infrastructure of multi-core or many-core clusters is a challenge for developers without experience in this domain. Developers need to use several libraries such as MPI, OpenMP, and CUDA to efficiently use the hardware which may include additional accelerators such as GPUs. Also, performing low-level optimizations is required in order to reach high performance. One approach to overcome these issues is the concept of Algorithmic Skeletons. These are instances of typical patterns for parallel programming, such as map, fold, and zip, which can simply be composed by an application programmer without taking care of low-level programming aspects. We propose a domain-specific language called Musket that includes algorithmic skeletons as domain abstractions which seamlessly integrate with sequential code while aligning with the C++ programming language for fast learnability. For improved usability, the editing component validates the correctness of models and provides solution hints in the integrated development environment. From the naive program specification, automatic transformations are applied in order to optimize the code for parallel execution. Subsequently, low-level C++ programs are generated which are optimized for multi-core parallelism on a cluster infrastructure. We evaluate the language using benchmark models written in our DSL and compare the execution time and speedup achieved through model preprocessing and code generation. Our experimental results show that the performance of Musket programs can be significantly improved through intermediate optimizations. The DSL approach thus simplifies multi-core application development and enables performance optimizations through model transformations.