In this survey paper, we compare native double precision solvers with emulated- and mixed- precision solvers of linear systems of equations as they typically arise in finite element discretisations. The emulation utilises two single float numbers to achieve higher precision, while the mixed precision iterative refinement computes residuals and updates the solution vector in double precision but solves the residual systems in single precision. Both techniques have been known since the 1960s, but little attention has been devoted to their performance aspects. Motivated by changing paradigms in processor technology and the emergence of highly parallel devices with outstanding single float performance, we adapt the emulation and mixed precision techniques to coupled hardware configurations, where the parallel devices serve as scientific co-processors. The performance advantages are examined with respect to speedups over a native double precision implementation (time aspect) and reduced area requirements for a chip (space aspect). The paper begins with an overview of the theoretical background, algorithmic approaches and suitable hardware architectures. We then employ several conjugate gradient and multigrid solvers and study their behaviour for different parameter settings of the iterative refinement technique. Concrete speedup factors are evaluated on the coupled hardware configuration of a general-purpose CPU and a graphics processor. The dual performance aspect of potential area savings is assessed on a field programmable gate array. In the last part, we test the applicability of the proposed mixed precision schemes with ill-conditioned matrices. We conclude that the mixed precision approach works very well with the parallel co-processors gaining speedup factors of four to five, and area savings of three to four, while maintaining the same accuracy as a reference solver executing everything in double precision

The first part of this paper surveys co-processor approaches for commodity based clusters in general, not only with respect to raw performance, but also in view of their system integration and power consumption. We then extend previous work on a small GPU cluster by exploring the heterogeneous hardware approach for a large-scale system with up to 160 nodes. Starting with a conventional commodity based cluster we leverage the high bandwidth of graphics processing units (GPUs) to increase the overall system bandwidth that is the decisive performance factor in this scenario. Thus, even the addition of low-end, out of date GPUs leads to improvements in both performance-and power-related metrics.

This article explores the coupling of coarse and fine-grained parallelism for Finite Element simulations based on efficient parallel multigrid solvers. The focus lies on both system performance and a minimally invasive integration of hardware acceleration into an existing software package, requiring no changes to application code. Because of their excellent price performance ratio, we demonstrate the viability of our approach by using commodity graphics processors (GPUs) as efficient multigrid preconditioners. We address the issue of limited precision on GPUs by applying a mixed precision, iterative refinement technique. Other restrictions are also handled by a close interplay between the GPU and CPU. From a software perspective, we integrate the GPU solvers into the existing MPI-based Finite Element package by implementing the same interfaces as the CPU solvers, so that for the application programmer they are easily interchangeable. Our results show that we do not compromise any software functionality and gain speedups of two and more for large problems. Equipped with this additional option of hardware acceleration we compare different choices in increasing the performance of a conventional, commodity based cluster by increasing the number of nodes, replacement of nodes by a newer technology generation, and adding powerful graphics cards to the existing nodes.

Abstract-We present a time skewing algorithm that breaks the memory wall for certain iterative stencil computations. A stencil computation, even with constant weights, is a completely memory-bound algorithm. For example, for a large 3D domain of 500 3 doubles and 100 iterations on a quad-core Xeon X5482 3.2GHz system, a hand-vectorized and parallelized naive 7-point stencil implementation achieves only 1.4 GFLOPS because the system memory bandwidth limits the performance. Although many efforts have been undertaken to improve the performance of such nested loops, for large data sets they still lag far behind synthetic benchmark performance. The state-of-art automatic locality optimizer PluTo [1] achieves 3.7 GFLOPS for the above stencil, whereas a parallel benchmark executing the inner stencil computation directly on registers performs at 25.1 GFLOPS. In comparison, our algorithm achieves 13.0 GFLOPS (52% of the stencil peak benchmark).We present results for 2D and 3D domains in double precision including problems with gigabyte large data sets. The results are compared against hand-optimized naive schemes, PluTo, the stencil peak benchmark and results from literature. For constant stencils of slope one we break the dependence on the low system bandwidth and achieve at least 50% of the stencil peak, thus performing within a factor two of an ideal system with infinite bandwidth (the benchmark runs on registers without memory access). For large stencils and banded matrices the additional data transfers let the limitations of the system bandwidth come into play again, however, our algorithm still gains a large improvement over the other schemes.

The solution of large sparse linear systems arises in many applications, such as computational fluid dynamics and oil reservoir simulation. In realistic cases the matrices are often so large that they require large scale distributed parallel computing to obtain the solution of interest in a reasonable time. In this paper we discuss the design and implementation of the AmgX library, which provides drop-in GPU acceleration of distributed algebraic multigrid (AMG) and preconditioned iterative methods. The AmgX library implements both classical and aggregation-based AMG methods with different selector and interpolation strategies, along with a variety of smoothers and preconditioners, including block-Jacobi, Gauss-Seidel, and incomplete-LU factorization. The library contains many of the standard and flexible preconditioned Krylov subspace iterative methods, which can be combined with any of the available multigrid methods or simpler preconditioners. The parallelism in the aggregation scheme exploits parallel graph matching techniques, while the smoothers and preconditioners often rely on parallel graph coloring algorithms. The AMG algorithm implemented in the AmgX library achieves 2-5× speedup on a single GPU against a competitive implementation on the CPU. As will be shown in the numerical experiments section, both setup and solve phases scale well across multiple nodes, sustaining this performance advantage.