A parallel iterative solver for large sparse linear systems enhanced with randomization and GPU accelerator, and its resilience to soft errors
Aygul Jamal · HAL (Le Centre pour la Communication Scientifique Directe) · 2017
In this PhD thesis, we address three challenges faced by linear algebra solvers in the perspective of future exascale systems: accelerating convergence using innovative techniques at the algorithm level, taking advantage of GPU (Graphics Processing Units) accelerators to enhance the performance of computations on hybrid CPU/GPU systems, evaluating the impact of errors in the context of an increasing level of parallelism in supercomputers. We are interested in studying methods that enable us to accelerate convergence and execution time of iterative solvers for large sparse linear systems. The solver specifically considered in this work is the parallel Algebraic Recursive Multilevel Solver (pARMS), which is a distributed-memory parallel solver based on Krylov subspace methods.First we integrate a randomization technique referred to as Random Butterfly Transformations (RBT) that has been successfully applied to remove the cost of pivoting in the solution of dense linear systems. Our objective is to apply this method in the ARMS preconditioner to solve more efficiently the last Schur complement system in the application of the recursive multilevel process in pARMS. The experimental results show an improvement of the convergence and the accuracy. Due to memory concerns for some test problems, we also propose to use a sparse variant of RBT followed by a sparse direct solver (SuperLU), resulting in an improvement of the execution time.Then we explain how a non intrusive approach can be applied to implement GPU computing into the pARMS solver, more especially for the local preconditioning phase that represents a significant part of the time to compute the solution. We compare the CPU-only and hybrid CPU/GPU variant of the solver on several test problems coming from physical applications. The performance results of the hybrid CPU/GPU solver using the ARMS preconditioning combined with RBT, or the ILU(0) preconditioning, show a performance gain of up to 30% on the test problems considered in our experiments.Finally we study the effect of soft fault errors on the convergence of the commonly used flexible GMRES (FGMRES) algorithm which is also used to solve the preconditioned system in pARMS. The test problem in our experiments is an elliptical PDE problem on a regular grid. We consider two types of preconditioners: an incomplete LU factorization with dual threshold (ILUT), and the ARMS preconditioner combined with RBT randomization. We consider two soft fault error modeling approaches where we perturb the matrix-vector multiplication and the application of the preconditioner, and we compare their potential impact on the convergence of the solver.