Exploring New Architectures in Accelerating CFD for Air Force Applications

Abstract

Abstract-1 Computational Fluid Dynamics (CFD) is an active field of research where the development of faster and more accurate methods is linked to the continuous demand for ever higher computational power. And indeed, for at least two decades, high-performance computing (HPC) programmers have taken for granted that each successive generation of microprocessors would, either immediately or after minor adjustments, make their software run substantially faster. But recent microprocessor design trends including the introduction of multi/many-core designs and the increasingly popular use in HPC of accelerators such as General Purpose Graphics Processing Units (GPGPU) and Field Programmable Gate Arrays (FPGAs), present an unprecedented challenge, namely how to update and enhance the existing large CFD software infrastructure to efficiently use these new architectures. In this paper we address some main issues in this transition and present ideas on using the new architectures to accelerate CFD applications that are of interest to the Air Force. We consider not only multi/many-core but also special purpose (e.g. GPUs) and reconfigurable computing (e.g. FPGAs) architectures. Moreover, we demonstrate benefits of using hybrid combinations where the strengths of each platform can be used to better map algorithm requirements and underlying architecture.

Towards dense linear algebra for hybrid GPU accelerated manycore systems

“…In May, Volkov and Demmel [35] described LU, QR, and Cholesky factorizations running at up to 180 GFlop/s in single precision (with QR a little bit more). The first results on a pre-released next-generation G90 NVIDIA card were presented at UGC2008 in May, where Dongarra et al [11] reported Cholesky running at up to 327 GFlop/s in single precision. Using again the newest generation card, in this paper, we describe an LU algorithm running at up to 388 GFlop/s in single precision and 99.4 Gflop/s in double precision.…”

“…This is illustrated for Cholesky factorization (so called left-looking version) in Fig. 2 (the case reported in [11]). The matrix to be factorized is allocated on the GPU memory and the code is as in LAPACK with BLAS calls replaced by CUBLAS, which represents the first idea from the list above.…”

“…Namely, while developing LAPACK-style onesided matrix factorizations for GPUs, several groups [2,4,36] observed that the panel factorizations are often faster on the CPU than on the GPU (the approach described in Section 2.2), which led to the development of highly efficient factorizations [11,29,34] where a single CPU core is used to control the GPU and to factor the panels. In contrast to this previous work, the importance here is on developing algorithms that will efficiently use both a multicore and a GPU.…”

The TheLMA project: A thermal lattice Boltzmann solver for the GPU

“…Their use in general-purpose computations [24], and more specifically in CFD [5], is promising. Successful attempts were made to implement LBM solvers on the GPU [6].…”

A Note on Auto-tuning GEMM for GPUs

“…Nevertheless, this approach can lead to high performance, but only after some modifications and for routines that map well on the GPU, like Cholesky (e.g. Dongarra et al [8] report up to 327 GFlop/s in single precision on a pre-released at the time NVIDIA T10P). Naturally, previous attempts to wrap some of the work needed in transitions like this in frameworks, have also failed to produce convincing results.…”