[Paper Review] IRPF90: a programming environment for high performance computing

Anthony Scemama|ArXiv.org|Sep 28, 2009

Parallel Computing and Optimization Techniques3 citations

TL;DR

IRPF90 is a Fortran 90 pre-processor that enables high-performance computing through a functional-style programming paradigm, automatically managing computational dependencies via implicit {needs/needed by} relationships. It reduces coding complexity, improves maintainability, and boosts performance—demonstrated by a 1.6× speedup in a quantum Monte Carlo application due to optimized memory access and elimination of redundant computations.

ABSTRACT

IRPF90 is a Fortran programming environment which helps the development of large Fortran codes. In Fortran programs, the programmer has to focus on the order of the instructions: before using a variable, the programmer has to be sure that it has already been computed in all possible situations. For large codes, it is common source of error. In IRPF90 most of the order of instructions is handled by the pre-processor, and an automatic mechanism guarantees that every entity is built before being used. This mechanism relies on the {needs/needed by} relations between the entities, which are built automatically. Codes written with IRPF90 execute often faster than Fortran programs, are faster to write and easier to maintain.

Motivation & Objective

• Address the complexity and maintainability challenges in large-scale Fortran HPC applications due to imperative programming's rigid control flow.
• Reduce programmer burden by automating the ordering of computations through implicit dependency tracking.
• Improve code efficiency and reusability by aligning programming structure with mathematical formulations using functional programming principles.
• Enable seamless integration with standard HPC libraries and Fortran 90 compilers without sacrificing performance.
• Support advanced features like checkpointing and variable declaration flexibility to enhance developer productivity.

Proposed method

• Use a pre-processor based on Python to translate IRPF90 source files into standard Fortran 90 code.
• Model computational entities as nodes in a dependency tree using implicit {needs/needed by} relationships derived from function calls.
• Automatically reorder computations via depth-first traversal of the dependency tree to ensure all required entities are computed before use.
• Allow local variable declarations anywhere in functions or subroutines, with the pre-processor relocating them to the beginning for standard Fortran compatibility.
• Introduce keywords like IRP_WRITE and IRP_READ to enable checkpointing by persisting computed entity values to disk.
• Use the PROVIDE keyword to explicitly define dependency relationships, especially useful in iterative processes or for performance tuning.

Experimental results

Research questions

• RQ1 Can a pre-processor that manages computational dependencies implicitly improve the maintainability and correctness of large Fortran HPC codes?
• RQ2 To what extent does implicit dependency management reduce redundant computations and improve performance in scientific applications?
• RQ3 How does the functional-style programming model in IRPF90 compare to traditional imperative coding in terms of developer productivity and code clarity?
• RQ4 What is the runtime performance overhead of the IRPF90 pre-processing and dependency management compared to hand-optimized Fortran code?
• RQ5 Can IRPF90 support advanced HPC features like checkpointing and modular code reuse without compromising portability or efficiency?

Key findings

• The IRPF90-generated code achieved a 1.60× speedup over the original Fortran implementation in a quantum Monte Carlo simulation, with CPU time reduced to 62% of the original.
• The performance gain was primarily due to reduced redundant computation and improved temporal locality, increasing the average instructions per cycle from 1.47 to 1.81.
• The pre-processor’s automatic dependency resolution eliminated the need for manual ordering, reducing code complexity and the risk of logic errors in large programs.
• Variable declarations can now be placed contextually within functions, improving code readability, with the pre-processor relocating them to the correct position automatically.
• The IRPF90 pre-processor generates standard Fortran 90 code fully compatible with HPC libraries such as BLAS, LAPACK, and MPI, ensuring portability and interoperability.
• Checkpointing functionality was added transparently via IRP_WRITE and IRP_READ statements, enabling persistent state saving and reloading without modifying the core logic.