Using GPUs to compute large out-of-card FFTs

Gu, Liang; Siegel, Jakob; Li, Xiaoming

doi:10.1145/1995896.1995937

Cited by 20 publications

(20 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As stated in the introduction, and as illustrated in the above figure, our performance is substantially better than those reported in the most recent related works of [1] and [11].…”

Section: A the Case Of Three Periodic Boundary Conditionssupporting

confidence: 38%

“…They reported a performance of around 50 GFLOPS on four nodes, somewhat lower than our performance on a single node with a Tesla C1060 (in fact, our performance number is an underestimate since it does not include all the components of our Poisson solver). Another recent work is reported by Gu et al [11], which tries to optimize both CPU-GPU data transfer and GPU computations for 1D, 2D, and 3D FFTs. In particular, they develop a blocked buffered technique for 1D FFTs which achieves a high bandwidth on the CPU-GPU data channel.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

High Performance FFT Based Poisson Solver on a CPU-GPU Heterogeneous Platform

JáJá

2013

2013 IEEE 27th International Symposium on Parallel and Distributed Processing

View full text Add to dashboard Cite

Abstract-We develop an optimized FFT based Poisson solver on a CPU-GPU heterogeneous platform for the case when the input is too large to fit on the GPU global memory. The solver involves memory bound computations such as 3D FFT in which the large 3D data may have to be transferred over the PCIe bus several times during the computation. We develop a new strategy to decompose and allocate the computation between the GPU and the CPU such that the 3D data is transferred only once to the device memory, and the executions of the GPU kernels are almost completely overlapped with the PCI data transfer. We were able to achieve significantly better performance than what has been reported in previous related work, including over 50 GFLOPS for the three periodic boundary conditions, and over 40 GFLOPS for the two periodic, one Neumann boundary conditions. The PCIe bus bandwidth achieved is over 5GB/s, which is close to the best possible on our platform. For all the cases tested, the single 3D PCIe transfer time, which constitutes a lower bound on what is possible on our platform, takes almost 70% of the total execution time of the Poisson solver.

show abstract

“…As stated in the introduction, and as illustrated in the above figure, our performance is substantially better than those reported in the most recent related works of [1] and [11].…”

Section: A the Case Of Three Periodic Boundary Conditionssupporting

confidence: 38%

Section: Introductionmentioning

confidence: 99%

High Performance FFT Based Poisson Solver on a CPU-GPU Heterogeneous Platform

JáJá

2013

2013 IEEE 27th International Symposium on Parallel and Distributed Processing

View full text Add to dashboard Cite

show abstract

“…To demonstrate the improvement of our PCI bandwidth, we used the same subarray test as Gu's work [6], where there are C regular subarrays of length W each.…”

Section: Pci Bandwidth Evaluationmentioning

confidence: 99%

A hybrid GPU/CPU FFT library for large FFT problems

Chen

2013

2013 IEEE 32nd International Performance Computing and Communications Conference (IPCCC)

Self Cite

View full text Add to dashboard Cite

“…Gu et al [12] propose a number of techniques for performing large FFTs when the data is maintained out of a single devices memory space.…”

Section: Related Workmentioning

confidence: 99%

Implementation of 3D FFTs Across Multiple GPUs in Shared Memory Environments

Nandapalan

Jaroš

Rendell

et al. 2012

2012 13th International Conference on Parallel and Distributed Computing, Applications and Technologies

View full text Add to dashboard Cite

Abstract-In this paper, a novel implementation of the distributed 3D Fast Fourier Transform (FFT) on a multi-GPU platform using CUDA is presented. The 3D FFT is the core of many simulation methods, thus its fast calculation is critical. The main bottleneck of the distributed 3D FFT is the global data exchange which must be performed. The latest version of CUDA introduces direct GPU-to-GPU transfers using a Unified Virtual Address space (UVA) that provides new possibilities for optimising the communication part of the FFT. Here, we propose different implementations of the distributed 3D FFT, investigate their behaviour, and compare their performance with the single GPU CUFFT and CPU-based FFTW libraries.In particular, we demonstrate the advantage of direct GPU-to-GPU transfers over data exchanges via host main memory. Our preliminary results show that running the distributed 3D FFT with four GPUs can bring a 12% speedup over the single node (CUFFT) while also enabling the calculation of 3D FFTs of larger datasets. Replacing the global data exchange via shared memory with direct GPU-to-GPU transfers reduces the execution time by up to 49%. This clearly shows that direct GPU-to-GPU transfers are the key factor in obtaining good performance on multi-GPU systems.

show abstract

Using GPUs to compute large out-of-card FFTs

Cited by 20 publications

References 12 publications

High Performance FFT Based Poisson Solver on a CPU-GPU Heterogeneous Platform

High Performance FFT Based Poisson Solver on a CPU-GPU Heterogeneous Platform

A hybrid GPU/CPU FFT library for large FFT problems

Implementation of 3D FFTs Across Multiple GPUs in Shared Memory Environments

Contact Info

Product

Resources

About