2decomp.org

One-line overview

2decomp.org is an open-source project focused on high-performance computing (HPC). It provides a Fortran library for “2D pencil decomposition” along with a companion distributed Fast Fourier Transform (FFT) library. Maintained by the UK academic and research community, it is primarily aimed at scientists and engineers running large-scale parallel workloads on supercomputers. Users choose it because it can efficiently decompose 3D grid data across thousands or even tens of thousands of CPU cores, making it a critical low-level tool for research workloads that require heavy FFT usage, such as meteorology, fluid dynamics, and molecular dynamics.

Business details

2decomp.org is not a commercial company, but an open-source software project backed by HPC teams at UK universities and research institutions. Its history dates back to the early 2000s, when the parallel scale of supercomputers began to grow rapidly and traditional one-dimensional or two-dimensional data decomposition methods started hitting communication-overhead bottlenecks. The 2D pencil decomposition technique emerged in response: it slices 3D data into “pencil-shaped” subdomains across a two-dimensional plane, significantly reducing cross-node communication latency. The libraries provided by the project have been widely integrated into well-known open-source CFD and scientific computing frameworks such as OpenFOAM, CAF, and Nek5000. In terms of industry standing, while it is not a commercial product, it has become one of the “standard” tools in academic HPC as one of the few open-source libraries specifically optimized for distributed FFT workloads. Its users are mainly university labs, national supercomputing centers, and research institutions in fields such as energy and aerospace.

Who it’s for

This project is best suited to the following types of users: first, research teams or graduate students who are writing or maintaining large-scale parallel numerical simulation codes in Fortran and need to distribute 3D data across thousands of processors; second, CFD (computational fluid dynamics) or climate model developers whose core algorithms depend on 3D FFTs or Poisson solvers; and third, maintainers of established Fortran projects who want to make their code run on next-generation exascale supercomputers without reinventing the wheel. Less suitable use cases include individual developers doing small-scale prototyping, non-Fortran users such as Python or C++ developers, and enterprise projects that require commercial technical support or SLA guarantees—because the project is entirely community-maintained and has no paid customer service.

Key features and highlights

• 2D pencil decomposition algorithm: The core selling point. It partitions 3D grids along two dimensions into pencil-shaped subdomains, significantly reducing MPI communication volume at extreme scale and offering better scalability than traditional 1D decomposition.
• Distributed FFT library: Implements parallel FFTs based on 2D decomposition, supports multiple data layouts such as complex-to-complex and real-to-complex, and can call FFTW or MKL directly as the underlying transform engine.
• Pure Fortran interface: Written entirely in Fortran 90/95, making it easy to integrate with existing Fortran code without language bindings or additional wrappers.
• Open source with an active community: The code is hosted on GitHub under a BSD or similarly permissive license, allowing free modification and commercial use; the mailing list and GitHub Issues tend to receive reasonably quick responses.
• Compatible with mainstream frameworks: It has been embedded as a core component in software such as OpenFOAM and CAF, so users can call its API directly without rewriting communication logic.
• Lightweight with minimal dependencies: No hard dependencies beyond MPI and FFTW, making it easy to compile and deploy on a wide range of Linux clusters.

Pricing analysis

2decomp.org is completely open source and free, with no plans, monthly fees, or annual fees. Users simply download the source code from GitHub or SourceForge and compile it themselves. From a cost perspective, it sits in the “zero-price” tier and offers excellent value—provided the user is capable of configuring MPI and a Fortran compiler in a supercomputing environment. The hidden cost is mainly labor: the build process may run into incompatibilities with specific system library versions and require manual debugging. In addition, deep customization or performance tuning usually requires developers with HPC experience. Compared with commercial alternatives such as Intel MKL’s cluster FFT module, which provides official support but requires expensive licensing, 2decomp.org is completely free—but offers no formal technical support.

How users in China can use it

According to official information, the code for 2decomp.org is hosted on GitHub. Direct downloads from mainland China are generally acceptable, though they may occasionally be interrupted by network instability. It is recommended to obtain source packages through domestic mirror sites such as Tsinghua University TUNA or USTC mirrors, or to use a Git proxy for acceleration. Compilation and execution take place entirely on a local supercomputing cluster or server, so no VPN or special network access is required. Since the software is free, there is no need to worry about foreign-currency payments. As for invoicing, because this is an open-source community project, it cannot provide Chinese tax invoices. If an enterprise user needs compliant reimbursement, it would have to work with a third-party HPC consulting company for integration services and invoicing. In terms of domestic alternatives, the Institute of Software at the Chinese Academy of Sciences has released a parallel FFT library with similar functionality, but it is far less widely adopted than 2decomp.org.

Pros and cons

Pros

• ✅ Completely free and open source, with no licensing restrictions
• ✅ Optimized for extreme-scale parallel workloads with tens of thousands of cores, offering excellent scalability
• ✅ Seamless fit with the Fortran ecosystem and a relatively gentle learning curve
• ✅ Solid community documentation and academic papers, with transparent technical principles
• ✅ Lightweight, with no unnecessary dependencies

Cons

• ❌ Fortran-only; no C/C++ or Python bindings provided
• ❌ No commercial support; bugs must be investigated by users themselves or left for the community to fix
• ❌ Domestic mirrors in China are unofficial, so users must handle source-code download issues themselves
• ❌ Cannot provide Chinese invoices, which may complicate enterprise procurement processes
• ❌ Narrow scope: focused only on 2D decomposition and FFTs, without a broader parallel computing toolkit

Comparison with similar products

• FFTW (MPI version): Also open source, but FFTW’s MPI interface uses 1D decomposition, which is less communication-efficient at extreme scale than 2decomp.org’s 2D pencil approach. FFTW supports C/Fortran and is more broadly applicable, but has a lower scalability ceiling.
• Intel MKL cluster FFT: Commercial software, deeply optimized for Intel hardware and backed by official technical support, but it requires a paid license and is limited to Intel platforms. It is not ideal for supercomputing users running AMD or ARM processors.
• P3DFFT: A similar open-source library that also implements 2D pencil decomposition, but it is primarily written in C while offering a Fortran interface. 2decomp.org is more established in the Fortran community, with richer documentation and more user case studies.

Final recommendation

2decomp.org is best suited to research teams that already use Fortran for large-scale parallel code and need to handle 3D FFTs or Poisson problems. If your project runs on a supercomputer with thousands of cores or more and you want to avoid implementing complex MPI communication patterns yourself, this library is close to a must-have. Conversely, if you use Python or C++, run on relatively small clusters of a few hundred cores or less, or require commercial-grade after-sales support, it may not be a good fit. The best approach is to download the source code, compile and test it on your local cluster, and then decide whether to integrate it into your core codebase—since it is completely free, the trial cost is effectively zero.