Parallel-in-time high-order multiderivative IMEX solvers

Jochen Schütz,David C. Seal,Jonas Zeifang

DOI: https://doi.org/10.48550/arXiv.2101.07846

2021-01-20

Abstract:In this work, we present a novel class of parallelizable high-order time integration schemes for the approximate solution of additive ODEs. The methods achieve high order through a combination of a suitable quadrature formula involving multiple derivatives of the ODE's right-hand side and a predictor-corrector ansatz. The latter approach is designed in such a way that parallelism in time is made possible. We present thorough analysis as well as numerical results that showcase scaling opportunities of methods from this class of solvers.

Numerical Analysis

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to develop a new type of parallel high - order time - integration scheme for solving additive ordinary differential equations (ODEs). Specifically, the authors propose a new class of parallel high - order implicit - explicit multi - derivative methods (IMEX solvers), aiming to improve the accuracy and efficiency of time integration and enable time - parallel computing on distributed or multi - core computer architectures.

### Main Problems and Background

1. **Problem Description**:
- The paper focuses on the numerical approximation of nonlinear differential equations of the form \( w'(t)=\Phi(w)\equiv\Phi_I(w)+\Phi_E(w), t\in(0, T_{\text{end}}) \).
- Here, \( \Phi \) is split into two parts: the stiff term \( \Phi_I \) and the non - stiff term \( \Phi_E \), to handle solutions at different scales.

2. **Existing Challenges**:
- Traditional methods such as Runge - Kutta and Adams face challenges in terms of efficiency and stability when dealing with complex ODEs.
- Time parallelization is crucial for efficient computing in modern multi - processor environments, but traditional time - integration methods are difficult to parallelize directly.

3. **Objectives**:
- Develop a high - order implicit - explicit multi - derivative method, combined with a prediction - correction strategy, to increase the overall accuracy.
- The designed method should be able to fully utilize time parallelism, thereby achieving better performance in multi - core or distributed computing environments.

### Method Innovation Points

1. **Multi - Derivative Runge - Kutta Method**:
- Use the second - order derivative to construct a numerical method, considering not only \( \Phi_E \) and \( \Phi_I \) but also higher - order time derivatives.
- This method can increase the order of consistency without increasing the number of stages or steps.

2. **Prediction - Correction Strategy**:
- Calculate a predicted value and multiple corrected values within each time step, gradually improving the accuracy.
- The correction step increases the accuracy order by one each time until the maximum accuracy determined by the underlying quadrature rule is reached.

3. **Time Parallelization**:
- Through careful design of dependencies, the prediction and correction steps can be executed in parallel.
- Similar to the concept of pipelining, information and data streams are directed so that correction iterations can be carried out on different processors.

### Application Areas

This method is applicable to solving ODEs in various scientific and engineering fields, such as meteorology, geophysics, aerodynamics, molecular dynamics, etc. In particular, when the ODEs are derived from the discretization of singularly perturbed partial differential equations, this type of method is especially effective.

### Summary

The core of the paper is to propose a novel time - parallel high - order multi - derivative IMEX solver, which solves the shortcomings of traditional methods in high - precision and parallel computing, and provides a more efficient tool for large - scale scientific computing.