Abstract
Solving Caputo fractional reaction-diffusion equations is computationally demanding due to the non-local nature of fractional derivatives. This paper presents an efficient parallel algorithm utilizing an implicit finite-difference method. The approach features a novel parallel solver for linear tridiagonal equations based on a reduced system elimination method. Experimental results demonstrate high accuracy and significant performance gains, achieving a 14.55x speedup with 16 parallel processes on Intel Xeon E5-2670 CPUs.
Introduction
Fractional Differential Equations (FDEs) are essential for modeling systems with "memory" or non-local behavior, such as anomalous diffusion in complex physical or financial environments. However, numerical simulation of FDEs is challenging because fractional derivatives depend on the function's entire history. While implicit finite-difference methods offer stability, they necessitate solving large linear systems at every time step, which is computationally prohibitive for serial processors. The challenge lies in parallelizing these solvers—specifically the tridiagonal matrix systems—without losing accuracy or incurring excessive communication overhead.
Methodology
The proposed parallel framework employs a divide-and-conquer strategy using the Message Passing Interface (MPI). The core of the algorithm is a specialized tridiagonal solver designed to handle the diagonally dominant matrices resulting from the discretization of the Caputo fractional equation.
Parallel Tridiagonal Solver
The main matrix is partitioned into $p$ blocks corresponding to the number of processors. A new elimination technique transforms these blocks into a "reduced system," allowing processors to compute local solutions independently. Communication is restricted to small amounts of boundary data exchanged with a master process to finalize the global solution.
Optimization Strategies
Efficiency is further enhanced by recognizing that the tridiagonal matrix $A$ remains constant across time iterations. By recording the initial matrix transformation, subsequent steps only require processing the changing right-hand side vector, significantly reducing redundant computations.
Results and Discussion
The algorithm was evaluated using a benchmark analytic problem ($u(x,t) = x(2-x)t^2$) and synthetic scaling tests with grid sizes ranging from $65 \times 65$ to $2,049 \times 2,049$.
Performance and Scalability
On a dual eight-core Intel Xeon E5-2670 setup, the implementation achieved a 14.55x speedup with 16 processes, representing a scaling efficiency of 91%. Notably, super-linear speedups were observed at low processor counts (2 and 4 processes), attributed to improved data locality and better CPU cache utilization as problem chunks became smaller.
Accuracy
The parallelization method maintains high precision, with a negligible difference ($4.22 \times 10^{-15}$) compared to the serial version. The maximum absolute error against the analytic solution was approximately $7.47 \times 10^{-4}$ for a $65 \times 65$ grid, decreasing consistently as resolution increased.
Conclusion
This research provides a robust parallel solution for Caputo fractional reaction-diffusion equations. By minimizing communication overhead and optimizing matrix operations, the algorithm enables large-scale simulations that were previously impractical. Future work will explore GPGPU acceleration and extending the elimination method to non-linear fractional equations.
Summary updated