Adaptively restarted block Krylov subspace methods with low-synchronization skeletons

TL;DR

This paper introduces adaptively restarted block Krylov subspace methods with low-synchronization skeletons, enhancing scalability and stability for high-performance computing applications like QR factorization on exascale systems.

Contribution

It transforms low-synchronization block Gram-Schmidt variants into block Arnoldi methods and develops an adaptive restarting heuristic to improve stability and performance.

Findings

Improved scalability of Krylov methods on exascale systems.

Enhanced stability with adaptive restarting heuristic.

Benchmarking shows competitive performance and accuracy.

Abstract

With the recent realization of exascale performace by Oak Ridge National Laboratory's Frontier supercomputer, reducing communication in kernels like QR factorization has become even more imperative. Low-synchronization Gram-Schmidt methods, first introduced in [K. \'{S}wirydowicz, J. Langou, S. Ananthan, U. Yang, and S. Thomas, Low Synchronization Gram-Schmidt and Generalized Minimum Residual Algorithms, Numer. Lin. Alg. Appl., Vol. 28(2), e2343, 2020], have been shown to improve the scalability of the Arnoldi method in high-performance distributed computing. Block versions of low-synchronization Gram-Schmidt show further potential for speeding up algorithms, as column-batching allows for maximizing cache usage with matrix-matrix operations. In this work, low-synchronization block Gram-Schmidt variants from [E. Carson, K. Lund, M. Rozlo\v{z}n\'{i}k, and S. Thomas, Block Gram-Schmidt algorithms and their stability properties, Lin. Alg. Appl., 638, pp. 150--195, 2022] are transformed into block Arnoldi variants for use in block full orthogonalization methods (BFOM) and block generalized minimal residual methods (BGMRES). An adaptive restarting heuristic is developed to handle instabilities that arise with the increasing condition number of the Krylov basis. The performance, accuracy, and stability of these methods are assessed via a flexible benchmarking tool written in MATLAB. The modularity of the tool additionally permits generalized block inner products, like the global inner product.