MM-AutoSolver: A multimodal machine learning method for the auto-selection of iterative solvers and preconditioners

The solution of large-scale sparse linear systems of the form $Ax=b$ is an important research problem in the field of High-performance Computing (HPC). With the increasing scale of these systems and the development of both HPC software and hardware, iterative solvers along with appropriate preconditioners have become mainstream methods for efficiently solving these sparse linear systems that arise from real-world HPC applications. Among abundant combinations of iterative solvers and preconditioners, the automatic selection of the optimal one has become a vital problem for accelerating the solution of these sparse linear systems. Previous work has utilized machine learning or deep learning algorithms to tackle this problem, but fails to abstract and exploit sufficient features from sparse linear systems, thus unable to obtain satisfactory results. In this work, we propose to address the automatic selection of the optimal combination of iterative solvers and preconditioners through the powerful multimodal machine learning framework, in which features of different modalities can be fully extracted and utilized to improve the results. Based on the multimodal machine learning framework, we put forward a multimodal machine learning model called MM-AutoSolver for the auto-selection of the optimal combination for a given sparse linear system. The experimental results based on a new large-scale matrix collection showcase that the proposed MM-AutoSolver outperforms state-of-the-art methods in predictive performance and has the capability to significantly accelerate the solution of large-scale sparse linear systems in HPC applications.