Can Fortran’s ‘do concurrent’ Replace Directives for Accelerated Computing?

M., Stulajter, Miko; Caplan, Ronald M.; Linker, J. A.

doi:10.1007/978-3-030-97759-7_1

Cited by 7 publications

(8 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Due to the novelty of DC support for GPUs, there are only a few examples of its use in the literature (such as a spherical surface diffusion tool called DIFFUSE [9], and an implementation of the BabelStream benchmark [10]). There are additional works in progress, including a chemistry mini-app called CCSD(T) 4 , a CFD weather mini-app called MiniWeather 5 , a hydrodynamics mini-app called CloverLeaf 6 , and a conjugate gradient solver used in Solar physics called POT3D 7 .…”

Section: Related Workmentioning

confidence: 99%

“…It can run simulations containing over three hundred million grid cells [23], [24] and exhibits performance scaling to thousands of CPU cores or dozens of GPUs [22], [25]. The MAS code uses a logically rectangular non-uniform staggered spherical grid and finite-difference discretizations with a combination of explicit and implicit time-stepping methods 9 . The code is highly memory-bound, with its performance typically proportional to the hardware's memory bandwidth.…”

Section: The Mas Solar Mhd Modelmentioning

confidence: 99%

“…From our previous experience in porting OpenACC to DC [7], we learned that using manual data management yields better performance than relying on unified managed memory (UM) capabilities. Unified managed memory is an NVIDIA feature that will automatically page data to and from the GPU and CPU for GPU-accelerated codes 10 .…”

Section: A Code 1 [A]: Current Openacc Implementationmentioning

confidence: 99%

“…Recently, the use of standard language parallelism for accelerated computing has been gaining popularity. This includes C++ parallel algorithms [5], [6], Fortran's do concurrent [7], [8], and drop-in replacements for python's numpy 1 . These new language features have the potential to eliminate (or greatly limit) the need for external acceleration APIs, and could be more portable and 'futureproof'.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Acceleration of a production Solar MHD code with Fortran standard parallelism: From OpenACC to `do concurrent'

Caplan¹,

M.²,

Linker³

2023

Preprint

View full text Add to dashboard Cite

There is growing interest in using standard language constructs for accelerated computing, avoiding the need for (often vendor-specific) external APIs. These constructs hold the potential to be more portable and much more 'future-proof'. For Fortran codes, the current focus is on the do concurrent (DC) loop. While there have been some successful examples of GPU-acceleration using DC for benchmark and/or small codes, its widespread adoption will require demonstrations of its use in full-size applications. Here, we look at the current capabilities and performance of using DC in a production application called Magnetohydrodynamic Algorithm outside a Sphere (MAS). MAS is a state-of-the-art model for studying coronal and heliospheric dynamics, is over 70,000 lines long, and has previously been ported to GPUs using MPI+OpenACC. We attempt to eliminate as many of its OpenACC directives as possible in favor of DC. We show that using the NVIDIA nvfortran compiler's Fortran 202X preview implementation, unified managed memory, and modified MPI launch methods, we can achieve GPU acceleration across multiple GPUs without using a single OpenACC directive. However, doing so results in a slowdown between 1.25x and 3x. We discuss what future improvements are needed to avoid this loss, and show how we can still retain close to the original code's performance while reducing the number of OpenACC directives by a factor of five.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: The Mas Solar Mhd Modelmentioning

confidence: 99%

Section: A Code 1 [A]: Current Openacc Implementationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Acceleration of a production Solar MHD code with Fortran standard parallelism: From OpenACC to `do concurrent'

Caplan¹,

M.²,

Linker³

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…We have implemented options both to use device memory in DoConcurrent and to use managed memory in the other models, to allow apples-to-apples and oranges-to-oranges comparisons, but results with those options are not included. There is value in evaluating idiomatic versus mixed-mode programming in the context of more complex algorithms and applications, as has been done in [43].…”

Section: Gpu Experimentsmentioning

confidence: 99%

Benchmarking Fortran DO CONCURRENT on CPUs and GPUs Using BabelStream

Hammond¹,

Deakin

Cownie

et al. 2022

2022 IEEE/ACM International Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems

View full text Add to dashboard Cite

Fortran DO CONCURRENT has emerged as a new way to achieve parallel execution of loops on CPUs and GPUs. This paper studies the performance portability of this construct on a range of processors and compares it with the incumbent models: OpenMP, OpenACC and CUDA. To do this study fairly, we implemented the BabelStream memory bandwidth benchmark from scratch, entirely in modern Fortran, for all of the models considered, which include Fortran DO CONCURRENT, as well as two variants of OpenACC, four variants of OpenMP (2 CPU and 2 GPU), CUDA Fortran, and both loop-and array-based references. BabelStream Fortran matches the C++ implementation as closely as possible, and can be used to make language-based comparisons. This paper represents one of the first detailed studies of the performance of Fortran support on heterogeneous architectures; we include results for AArch64 and x86 64 CPUs as well as AMD, Intel and NVIDIA GPU platforms.

show abstract