TL;DR
This paper introduces the "Object-Oriented Pattern Language" (OPL), a comprehensive design pattern language aimed at structuring parallel software architecture. By merging two prior projects, OPL provides a hierarchical framework of patterns to guide engineers in building correct, efficient, and productive parallel applications.
What Problem Does This Paper Address?
The authors argue that parallel programming has reached a plateau, largely due to a lack of a common, structured approach to designing parallel software architectures. They posit that the key to advancing parallel programming lies in understanding the software people want to engineer and providing them with a robust design pattern language. This language should serve as a shared vocabulary and a set of building blocks for creating complex parallel systems, moving beyond ad-hoc solutions. The paper aims to present a "grand consensus pattern language" to stimulate discussion and achieve a meaningful consensus in the field.
Methods
The paper presents a hierarchical design pattern language, OPL, which is a merger of two previous projects: PLPP and OPL. The language is structured into five major categories of patterns, organized into layers that correspond to the phases of mapping a problem onto a parallel computation.
The top two categories, Structural Patterns and Computational Patterns, form the high-level architecture layer. Structural patterns describe the overall organization and interaction of computational elements, akin to "boxes and arrows" diagrams. Computational patterns, on the other hand, define the classes of computations that make up the application, essentially describing "computations occurring in the boxes." These two categories are tightly coupled, with architects often choosing a structural pattern and then considering the necessary computational patterns.
The remaining three layers focus specifically on parallel programming:
- Parallel Algorithm Strategy Patterns: These define high-level strategies for exploiting concurrency within a computation for parallel execution.
- Implementation Strategy Patterns: These describe the structures realized in source code to organize the program and common data structures for parallel programming.
- Parallel Execution Patterns: These are approaches embodied in runtime systems to support the execution of parallel programs.
Beneath these layers are Parallel Foundation Constructs, which are common "hooks" for connecting algorithms to programming language constructs, though these are not discussed in detail in this paper.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- The "Seminal Intellectual Challenge": The authors frame the creation of a comprehensive design pattern language for all software engineering, particularly parallel software, as the most significant intellectual challenge facing computer science today.
- Hierarchy as a Solution: The paper emphasizes a hierarchical approach to design patterns, mirroring the stages of problem-solving from high-level architecture down to execution. This allows for abstraction and modularity.
- Bridging the Gap: OPL aims to bridge the gap between abstract computational problems and their concrete parallel implementations by providing patterns at various levels of abstraction.
- Consensus is Key: The authors stress the importance of broad community engagement and debate to achieve a meaningful consensus on the proposed pattern language, moving beyond discussions among a select few.
- Interconnectedness of Patterns: The paper highlights how patterns across different layers are naturally coupled, with choices in one layer influencing or being influenced by choices in others (e.g., a "divide and conquer" strategy often implies a "fork-join" implementation).
Key Results
The paper presents a structured overview of the OPL design pattern language, categorizing patterns into five main groups: Structural, Computational, Parallel Algorithm Strategy, Implementation Strategy, and Parallel Execution Patterns. It provides brief descriptions for each pattern within these categories, illustrating the proposed framework for architecting parallel software. The primary "result" is the proposed OPL itself, serving as a foundation for further discussion and refinement.
Limitations and Future Work
The authors explicitly state that this paper is a "snapshot" and is "stripped of motivation, background, related-work or other expository elements" to stimulate debate. They acknowledge that the next step is to write detailed text for each pattern and refine them through workshops. They also note that a pattern language alone is not a programming language and that frameworks are needed to translate these designs into executable code. Future work will involve completing the pattern descriptions and developing these supporting frameworks. The paper also poses three critical questions for debate: the completeness of the pattern set, the correctness of the layer definitions, and the accuracy of the pattern descriptions.
Summary updated