TL;DR
This paper introduces PFD-PARALLEL, a novel, open-source software package designed for massively parallel multivariate partial fractioning. It tackles the computational bottleneck of simplifying complex rational functions in high-energy physics, particularly in scattering amplitude calculations, by combining advanced algorithms with a sophisticated parallelization framework.
What Problem Does This Paper Address?
In theoretical high-energy physics, especially when calculating scattering amplitudes at higher orders, researchers often encounter extremely complex rational functions. These functions, which arise from intricate calculations like multi-loop integrals and integration-by-parts (IBP) reductions, are computationally challenging to handle and simplify. Existing methods for partial fraction decomposition, while powerful, struggle with the sheer scale and complexity of these expressions, leading to significant computational time and memory requirements. The authors aim to overcome this by developing a highly efficient, parallelized tool that can handle these large-scale problems.
Methods
The core of PFD-PARALLEL is the implementation and combination of two key algorithms: an improved version of Leinartas' algorithm and the MULTIVARIATEAPART algorithm. The novelty lies in their integration within a large-scale parallel framework built upon the SINGULAR/GPI-SPACE system. This framework utilizes Petri nets to model the parallel workflow, enabling parallelism at two levels:
- Parallelism across different rational function coefficients: This allows the processing of multiple independent rational functions simultaneously.
- Parallelism for individual expressions: This tackles the complexity within a single, very large rational function by decomposing its sub-components in parallel.
The system is entirely based on open-source software, making it accessible and adaptable.
Datasets
The paper does not discuss datasets in the traditional machine learning sense. Instead, it uses real-world computational problems from high-energy physics as its "datasets" for evaluation. These include:
- IBP reduction coefficients for two-loop five-point massless Feynman integrals: Specifically, they analyze integrals with numerator degrees up to 5.
- Rational functions from two-loop leading color helicity amplitudes for Wγ+j production: These are complex expressions arising from LHC physics.
The paper provides details on the size of these input "datasets" (e.g., 19.6 GB for IBP coefficients) and the resulting compressed output.
Key Insights and Ideas
- Two-level parallelism is crucial: Simply parallelizing over different rational functions isn't enough for extremely complex cases. Parallelizing the decomposition within individual functions is essential for significant speedups.
- Petri nets provide a powerful modeling tool: The use of Petri nets within the SINGULAR/GPI-SPACE framework offers a structured and automated way to manage complex, multi-layered parallel workflows.
- Open-source is key: Building the entire framework on open-source components (SINGULAR, GPI-SPACE) ensures accessibility, maintainability, and extensibility.
- Significant data reduction is achievable: The PFD-PARALLEL package can reduce the size of complex rational function expressions by factors of over 100, making subsequent analysis and computation much more feasible.
Key Results
The paper demonstrates the effectiveness of PFD-PARALLEL through two main examples:
- IBP Reduction Coefficients: For a set of 47 top-sector integrals with ISPs up to degree 5, the original data size of 19.6 GB was reduced to 186 MB, achieving a compression ratio of 108. This was accomplished in about 22 hours and 15 minutes on 520 cores using the two-layer parallelism. The compression ratio increases with the numerator degree.
- Analytic Multi-Loop Scattering Amplitudes: For 452 rational functions from Wγ+j production, the input size of 1096.76 MB was reduced to 9.20 MB (indexed text output), with a compression ratio of 187.60. This computation took 5 hours and 34 minutes on 48 CPU cores.
The paper highlights that the improved Leinartas algorithm and MULTIVARIATEAPART algorithm, when implemented within this parallel framework, achieve similar performance in terms of output size and computation time for these examples.
Limitations and Future Work
- Scheduling of long-running jobs: The paper notes that while the framework achieves good resource utilization, the presence of a few very long-running coefficients can still dominate the overall runtime. A variation of the WaitAll net is proposed to address this by pre-sorting tasks by size.
- MultivariateApart parallelism: The paper mentions that the MULTIVARIATEAPART algorithm is not parallelized within the framework, affecting only sequential processing of individual functions.
- Future improvements: The authors suggest that the code is designed for easy integration of future algorithmic improvements. They also note that a version of the framework allowing individual core count specification per node is under development. The Petri net based parallelization methods are expected to be useful for broader applications beyond partial fractioning.
Summary updated