Hierarchical Jacobi Iteration for Structured Matrices on GPUs using Shared Memory

TL;DR

This paper introduces a hierarchical Jacobi iteration algorithm optimized for GPUs that leverages shared memory to perform multiple iterations within each cycle, significantly accelerating convergence for structured linear systems.

Contribution

The paper presents a novel GPU algorithm that uses shared memory and domain decomposition to perform multiple Jacobi iterations efficiently, reducing global memory accesses and speeding up convergence.

Findings

Achieved up to 8x speedup in 1D problems

Achieved nearly 6x speedup in 2D problems

Effective utilization of shared memory improves iterative solver performance

Abstract

High fidelity scientific simulations modeling physical phenomena typically require solving large linear systems of equations which result from discretization of a partial differential equation (PDE) by some numerical method. This step often takes a vast amount of computational time to complete, and therefore presents a bottleneck in simulation work. Solving these linear systems efficiently requires the use of massively parallel hardware with high computational throughput, as well as the development of algorithms which respect the memory hierarchy of these hardware architectures to achieve high memory bandwidth. In this paper, we present an algorithm to accelerate Jacobi iteration for solving structured problems on graphics processing units (GPUs) using a hierarchical approach in which multiple iterations are performed within on-chip shared memory every cycle. A domain decomposition style procedure is adopted in which the problem domain is partitioned into subdomains whose data is copied to the shared memory of each GPU block. Jacobi iterations are performed internally within each block's shared memory, avoiding the need to perform expensive global memory accesses every step. We test our algorithm on the linear systems arising from discretization of Poisson's equation in 1D and 2D, and observe speedup in convergence using our shared memory approach compared to a traditional Jacobi implementation which only uses global memory on the GPU. We observe a x8 speedup in convergence in the 1D problem and a nearly x6 speedup in the 2D case from the use of shared memory compared to a conventional GPU approach.