Accelerating the Solution of Sparse Linear Systems with GPUs

Friday November 9, 2018
Location: Hamerschlag Hall D210
Time: 10:00AM-11:00AM

Krylov solvers are one of the most commonly used methods for the solution of large, sparse linear systems. However, depending on the numerical properties of the system matrix these methods may converge slowly, or even diverge. Preconditioning is an effective method which aims to improve the numerical properties of the matrix, and thus, the convergence rate of the Krylov solver. As conventional ILU-based preconditioners offer only a limited degree of concurrency, in both the factorization as well as the triangular solve phases, they are often not able to fully utilize the highly-parallel accelerator hardware. In contrast, while usually offering a lower convergence rate improvement, block-Jacobi preconditioning can be expressed in terms of independent computations on small diagonal blocks, making it an ideal candidate for these types of machines.

In this talk, I will describe several approaches for efficient generation and application of block-Jacobi preconditioners on CUDA-enabled GPUs. All presented kernels heavily rely on the increased register count and warp shuffle instructions available in newer GPU hardware and are tuned specifically for small problem sizes arising in block-Jacobi preconditioning. As a result, these routines outperform their counterparts from both open-source, as well as proprietary libraries, by a large margin. Finally, I will give an overview of the current research pursued by our group, which aims to further improve the effectiveness of block-Jacobi preconditioning, by reducing its memory footprint, as well as its energy consumption.

Goran Flegar is a PhD candidate in the High Performance Computing & Architectures group at the Jaume I University. His research interests include sparse linear algebra, accelerator computing and software design. He holds a bachelor's degree in mathematics and a master's degree in computer science and mathematics from the University of Zagreb. He is also one of the main developers of Ginkgo, a modern C++ library for the iterative solution of sparse linear systems, via Krylov subspace methods, on multicore architectures.