Python accelerators for high-performance computing

Marowka, Ami

doi:10.1007/s11227-017-2213-5

Cited by 33 publications

(22 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…In [ 15 ] the same application is evaluated in Python/Numba and CUDA/Fortran. In [ 12 ] a detailed analysis of Numba is provided, but the authors do not provide a detailed comparison to C-CUDA. In [ 13 ] we compared the performance of C-CUDA and Numba-CUDA, but mainly focus on the performance of the compute kernels.…”

Section: Related Workmentioning

confidence: 99%

Implementation and Evaluation of CUDA-Unified Memory in Numba

Oden¹,

Saidi²

2021

Euro-Par 2020: Parallel Processing Workshops

View full text Add to dashboard Cite

Python as a programming language is increasingly gaining importance, especially in data science, scientific, and parallel programming. With the Numba-CUDA, it is even possible to program GPUs with Python using a CUDA like programming style. However, Numba is missing support for CUDA-unified memory, which can help to simplify programming even more and allows dynamic work distribution between GPUs and CPUs. In this work, we implement and evaluate the support for unified memory in Numba. As expected, the performance of unified memory is worse than using explicit data transfers, but can outperform the performance of the implicit methods provided by Numba. Additionally, using unified memory can help to reduce the Python interpreter overhead and therefore help to improve the performance of small problem sizes. The use of system-wide atomic can help to improve the work distribution between GPU and CPU, but when using more CPU threads the performance suffers under the Python global interpreter lock (GIL).

show abstract

Section: Related Workmentioning

confidence: 99%

Implementation and Evaluation of CUDA-Unified Memory in Numba

Oden¹,

Saidi²

2021

Euro-Par 2020: Parallel Processing Workshops

View full text Add to dashboard Cite

show abstract

“…For example, they allow resampling any time series in a given time span. We look forward to the progress of projects such as PyPy, Numba, or Grumpy, which aim to dilute the boundaries between high and low‐level programming languages, bringing together the strong abstraction of the former while keeping the high performance of the latter. We note that they still require further refinement and wider compatibility but the trend of high‐level languages with faster just‐in‐time compilers looks promising for the upcoming challenges of automatic traffic reporting and network management.…”

Section: Architecturementioning

confidence: 99%

On the design and performance evaluation of automatic traffic report generation systems with huge data volumes

Vega

Sierra

Moreno³

et al. 2018

Int J Network Mgmt

View full text Add to dashboard Cite

In this paper, we analyze the performance issues involved in the generation of automated traffic reports for large IT infrastructures. Such reports allow the IT manager to proactively detect possible abnormal situations and roll out the corresponding corrective actions. With the ever-increasing bandwidth of current networks, the design of automated traffic report generation systems is very challenging. In a first step, the huge volumes of collected traffic are transformed into enriched flow records obtained from diverse collectors and dissectors. Then, such flow records, along with time series obtained from the raw traffic, are further processed to produce a usable report. As will be shown, the data volume in flow records turns out to be very large as well and requires careful selection of the key performance indicators (KPIs) to be included in the report. In this regard, we discuss the use of high-level languages versus low-level approaches, in terms of speed and versatility. Furthermore, our design approach is targeted for rapid development in commodity hardware, which is essential to cost-effectively tackle demanding traffic analysis scenarios. Actually, the paper shows feasibility of delivering a large number of KPIs, as will be detailed later, for several TBytes of traffic per day using a commodity hardware architecture and high-level languages.

show abstract

“…The next set of papers concern language-based parallelization techniques. Marowka [6] investigates the efforts devoted by the Python community to improve the performance of this language in scientific applications. Mostly, he focuses on Numba, a much-promised solution that preserves the performance improvement in code ported to different target architectures.…”

mentioning

confidence: 99%

Special section on parallel programming

Marowka¹,

Stpiczyński

2018

J Supercomput

Self Cite

View full text Add to dashboard Cite

WLPP 2017 was a two-day workshop focusing on high-level programming for large-scale parallel systems and multicore processors, with special emphasis on component architectures and models. Its goal was to bring together researchers working in the areas of applications, computational models, language design, compilers, system architecture, and programming tools to discuss new developments in programming Clouds and parallel systems.Papers in this section cover the most important topics presented during the workshop. The first two deal with parallel programming models. Thoman et al. [4] provide an initial task-focused taxonomy for HPC technologies, which covers both programming interfaces and runtime mechanisms and discuss its usefulness by classifying state-of-the-art task-based environments that are used today. Posner and Fohry [7] propose a hybrid work stealing scheme, which combines the lifeline-based variant of distributed task pools with the node-internal load balancing implemented as an extension of the APGAS library for Java. The next set of papers concern language-based parallelization techniques. Marowka [6] investigates the efforts devoted by the Python community to improve the performance of this language in scientific applications. Mostly, he focuses on Numba, a much-promised solution that preserves the performance improvement in code ported to different target architectures. Stpiczyński [8] evaluates intrinsics, OpenMP, TBB and Cilk Plus as basic language-based tools for simple and efficient optimization of computational problems that need both task and data parallelization techniques. Blöcker and Hoffmann [2] propose a fully deterministic process calculus for parallel and distributed programming implemented in Haskell as a domain-specific language. Batko and Kuta [1] compare their new functional language Anemone with actor systems of Scala and Erlang.Finally, four papers present the use of parallel programming for solving real-world computational problems. Bylina [3] demonstrates the methodology for parallelization of finding stochastic bounds for Markov chains on multicore and manycore platforms using OpenMP and Cilk Plus. Tokura et al.[10] present an almost optimal columnwise prefix-sum algorithm for GPUs. Halver et al. [5] discuss the use of OpenCL as a language for implementation of classical molecular dynamics simulation useful for benchmarking different hardware platforms. Szustak [9] presents an innovative strategy for the data flow synchronization in stencil parallel computations on sharedmemory systems.Acknowledgements The guest editors of this special issue wish to thank the authors for their interesting contributions and the reviewers for useful suggestions. Finally, we would like to express our gratitude to Professor Hamid Arabnia (Editor-in-Chief of The Journal of Supercomputing) for the opportunity to edit this special issue and his great guidance.

show abstract

Python accelerators for high-performance computing

Cited by 33 publications

References 1 publication

Implementation and Evaluation of CUDA-Unified Memory in Numba

Implementation and Evaluation of CUDA-Unified Memory in Numba

On the design and performance evaluation of automatic traffic report generation systems with huge data volumes

Special section on parallel programming

Contact Info

Product

Resources

About