AIMES: Advanced Computation and I/O Methods for Earth-System Simulations

Julian M. Kunkel, Nabeeh Jumah, Anastasiia Novikova, Thomas Ludwig, Hisashi Yashiro, Naoya Maruyama, Mohamed Wahib, John Thuburn, Thomas Ludwig 0002Software for Exascale Computing 2020 · DOI · DBLP

Introduction

Earth System Models (ESMs) are currently navigating a "perfect storm" of computational challenges. As the scientific community moves into the CMIP6 era, data requirements are escalating into the petabyte range. Simultaneously, the underlying supercomputing hardware is shifting from homogeneous CPU clusters to complex, heterogeneous architectures involving GPUs and many-core processors.

Existing ESM codebases, often written in legacy Fortran, are difficult to optimize for these new architectures without exhaustive manual rewrites. Furthermore, modern global models are transitioning away from traditional rectangular grids—which suffer from mathematical singularities at the poles—toward icosahedral (geodesic) grids. While these grids offer more uniform global coverage, they introduce significant complexity in programming and optimization.

Triangular (a) and hexagonal (b) grid structures used in modern icosahedral Earth System Models.

The AIMES Approach

The AIMES project addresses these challenges by decoupling scientific formulation from hardware-specific optimization. This ensures "performance portability"—the ability for a single scientific code to run efficiently on various hardware types—without sacrificing researcher productivity. The project introduces two primary software solutions: GGDML and SCIL.

GGDML: Simplifying Grid Manipulation

The General Grid Definition and Manipulation Language (GGDML) is a lightweight Domain-Specific Language (DSL) designed as an extension to existing languages like Fortran or C. It abstracts the complexities of grid traversal and memory access. A Python-based source-to-source (S2S) translation tool processes this high-level code to generate optimized versions for specific backends, such as MPI+OpenMP or OpenACC.

SCIL: Automating I/O Optimization

To manage the data explosion, the Scientific Compression Library (SCIL) acts as a "meta-compressor." Instead of requiring scientists to manually tune complex compression algorithms (like ZFP or SZ), SCIL allows them to specify a tolerable error (e.g., "1% absolute error"). The library then automatically selects the optimal compression chain to meet these constraints while maximizing I/O throughput.

Key Insights and Results

The AIMES framework demonstrates that high-level abstractions can lead to significant gains in both developer efficiency and hardware performance:

  • Code Complexity: Rewriting computational kernels in GGDML resulted in an average 70% reduction in Lines of Code (LOC) compared to original Fortran implementations.
  • Performance Portability: GGDML-generated code achieved approximately 80% of theoretical peak memory throughput across diverse architectures, including Broadwell CPUs, P100 GPUs, and NEC Aurora vector engines.
  • I/O Acceleration: By utilizing lossy compression via SCIL, the team achieved I/O performance improvements of up to 13x on the Mistral supercomputer. This highlights that lossy compression is not just a storage-saving measure but a critical tool for overcoming I/O bottlenecks that stall simulations.
  • Kernel Fusion: The researchers found that merging separate computational steps (inter-kernel fusion) to keep data in the processor's cache resulted in a 30–50% performance increase for memory-bound models.

Conclusion and Future Directions

The AIMES project proves that the gap between scientific productivity and exascale performance can be bridged through intelligent abstraction. While the current implementation requires some manual porting of existing models, the results in code reduction and I/O speedup are compelling. Future work aims to expand GGDML to include I/O-relevant specifications, potentially creating a unified framework for optimizing both the computation and storage phases of Earth-system simulations.

Summary updated

BibTeX
@inproceedings{kunkel2020aimes,
  author = {Julian M. Kunkel and Nabeeh Jumah and Anastasiia Novikova and Thomas Ludwig and Hisashi Yashiro and Naoya Maruyama and Mohamed Wahib and John Thuburn and Julian Kunkel and Thomas Ludwig},
  title = {{AIMES: Advanced Computation and I/O Methods for Earth-System Simulations}},
  booktitle = {Software for Exascale Computing},
  year = {2020},
  doi = {10.1007/978-3-030-47956-5\_5},
}
Figure