Search

We present a new model called GPUFLOW for the modeling and simulation of geophysical flows. GPUFLOW, which is based on the cellular automaton paradigm, features a physical model for the thermal and rheological evolution of lava flows (including temperature-dependent emissivity and cooling by radiation and air convection), support for debris flows without thermal dependency, a parallel implementation on graphic processing units (GPUs), and a simpler and computationally more efficient solution to the grid bias problem. Here, we describe the physical–mathematical model implemented in GPUFLOW and estimate the influence of input data on the flow emplacement through different synthetic test cases. We also perform a validation using two real applications: a debris flow that occurred in July 2006 in the Dolomites (Italy) and the December 2018 lava flow from the Etna volcano. GPUFLOW’s reliability prediction is accomplished by fitting the simulation with the actual flow fields, obtaining average values between ~55% and 75%. more...

The exponentially increasing bioinformatics data raised a new problem: the computation time length. The amount of data that needs to be processed is not matched by an increase in hardware performance, so it burdens researchers on computation time, especially on drug-target interaction prediction, where the computational complexity is exponential. One of the focuses of high-performance computing research is the utilization of the graphics processing unit (GPU) to perform multiple computations in parallel. This study aims to see how well the GPU performs when used for deep learning problems to predict drug-target interactions. This study used the gold-standard data in drug-target interaction (DTI) and the coronavirus disease (COVID-19) dataset. The stages of this research are data acquisition, data preprocessing, model building, hyperparameter tuning, performance evaluation and COVID-19 dataset testing. The results of this study indicate that the use of GPU in deep learning models can speed up the training process by 100 times. In addition, the hyperparameter tuning process is also greatly helped by the presence of the GPU because it can make the process up to 55 times faster. When tested using the COVID-19 dataset, the model showed good performance with 76% accuracy, 74% F-measure and a speed-up value of 179. more...

3D models from the reconstruction of real-world objects have photorealistic details. This realistic detail in the visualization needs to be supported by a visualization environment that is able to represent realistic real-world conditions. However, the environmental light information on the reconstructed 3D object was lost in the reconstruction process. This is because the light obtained from scanned real-world objects for reconstruction can reduce the accuracy of the reconstruction results. This inaccuracy is caused by alterations in the light reflected from the captured object, causing the texture of the object difficult to reconstruct. In the visualization of the scanned object, the light that appears on the object is reformed virtually through ray tracing. This research utilizes the RT core technology on the Ampere graphics processor architecture introduced by Nvidia. With this technology, the ray tracing process can be carried out in real-time, enabling realistic and dynamic visualization. In this study, an application that is able to perform 3D visualization will be made by utilizing real-time ray tracing with an RT core on the graphics processing unit of the Ampere architecture. This application is able to blend virtual environments and 3D models that are reconstructed from the results of scanning real-world objects. We will demonstrate that this application can later become a 3D visualization solution for models from the reconstruction of real-world objects. more...

Several causes make brain cancer identification a challenging task for neurosurgeons during the surgical procedure. The surgeons' naked eye sometimes is not enough to accurately delineate the brain tumor location and extension due to its diffuse nature that infiltrates in the surrounding healthy tissue. For this reason, a support system that provides accurate cancer delimitation is essential in order to improve the surgery outcomes and hence the patient's quality of life. The brain cancer detection system developed as part of the “HypErspectraL Imaging Cancer Detection” (HELICoiD) European project meets this requirement exploiting a non-invasive technique suitable for medical diagnosis: the hyperspectral imaging (HSI). A crucial constraint that this system has to satisfy is providing a real-time response in order to not prolong the surgery. The large amount of data that characterizes the hyperspectral images, and the complex elaborations performed by the classification system make the High Performance Computing (HPC) systems essential to provide real-time processing. The most efficient implementation developed in this work, which exploits the Graphic Processing Unit (GPU) technology, is able to classify the biggest image of the database (worst case) in less than three seconds, largely satisfying the real-time constraint set to 1 minute for surgical procedures, becoming a potential solution to implement hyperspectral video processing in the near future. more...

Orientador: Luiz Otávio Saraiva Ferreira Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Mecânica Resumo: Os avanços na tecnologia das GPUs, tanto em hardware quanto em ferramentas de programação, tornaram-nas uma opção viável como processadores de uso geral, sendo adequadas para a realização de processamento paralelo com alta demanda de cálculos. Dentre os problemas que podem fazer uso das GPUs são as simulações que envolvem o fluxo de partículas como o Método dos Elementos Discretos (DEM). Assim, a proposta deste trabalho foi implementar em cluster de GPUs um simulador utilizando o Método dos Elementos Discretos. O simulador foi inicialmente validado para uma única GPU utilizando resultados experimentais disponíveis na literatura, onde foram possíveis obter resultados com erros menores do que 10%. Além disso, os tempos de processamento para uma única GPU foram comparados com outro simulador, também implementado em GPU, resultando em tempos de execução semelhantes aos reportados. Finalmente, o método foi expandido para o cluster de GPUs, utilizando uma abordagem híbrida (MPI + CUDA), e apresentou um ganho de desempenho adequado à medida que o número de GPUs foi aumentado Abstract: Advances in GPU technology, both in hardware and programming tools, have made them a viable option as general-purpose processors, suitable for performing parallel processing tasks with high computational demand. The Discrete Element Method (DEM) studies a class of problems that can make use of GPUs high computational power. Thus, the proposal of this work was to implement a simulator using the Discrete Element Method in a GPU cluster. The simulator was initially validated for a single GPU using experimental results available in the literature, where it was possible to obtain results with errors less than 10%. Also, processing times for a single GPU were compared with another simulator, also implemented in GPU, resulting in run times similar to those reported. Finally, the method was ported to the GPU cluster, using a hybrid approach (MPI + CUDA), and presented a suitable gain of performance as the number of GPUs was increased Mestrado Mecânica dos Sólidos e Projeto Mecânico Mestre em Engenharia Mecânica FAPESP 10/09717-7 more...

We present a CUDA-based parallel implementation on GPU architecture of a modified version of the Smoothed Particle Hydrodynamics (SPH) method. This modified formulation exploits a strategy based on the Taylor series expansion, which simultaneously improves the approximation of a function and its derivatives with respect to the standard formulation. The improvement in accuracy comes at the cost of an additional computational effort. The computational demand becomes increasingly crucial as problem size increases but can be addressed by employing fast summations in a parallel computational scheme. The experimental analysis showed that our parallel implementation significantly reduces the runtime, with speed-ups of up to 90,when compared to the CPU-based implementation. more...

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - CAPES Context: The popularization of Deep Learning and the Deep Inference spurred the development of tools for the implementation of Convolutional Neural Networks (CNNs) as specific libraries for the development of CNNs, the popularization also stimulated the development of GPUs with CNN acceleration capabilities. Developers and scientists working with CNNs need experimental scientific studies that point out which library is most suitable for a particular GPU microarchitecture. Objective: Compare the performance of the CNTK, PyTorch, TensorFlow1.15 and TensorFlow 2.2 libraries in different GPU microarchitectures (Kepler, Maxwell, Pascal and Turing) using LeNet-5 CNN and the MNIST dataset. Method: Initially, a systematic mapping was carried out as a way to identify and systematize the main benchmarks, then an experiment was carried out evaluating the impact of the Keras API on the performance of the libraries used in the comparative study and lastly, four experiments were carried out comparing the libraries, each experiment was performed in a computational environment accelerated by a GPU with a different microarchitecture. Results: For the systematic mapping, it was identified that the classic CNN architectures such as LeNet-5 and AlexNet are the most used as benchmarks, also showed that the most used data benchmarking sets are ImageNet, MNIST and CIFAR-10. The experiment on the impact of the Keras API showed that the API negatively impacts the performance of all the tested libraries. In the comparative study of the libraries’ performance, the PyTorch library presented the performance performance and the CNTK, TensorFlow 1.15 and TensorFlow 2.2 libraries alternated between the three shorter execution times. Conclusion: The results show that the PyTorch library presents a low level of use of GPU and uses a large amount of sec memory compared to other libraries that showed a higher performance. Contexto: A popularização da Deep Learning e da Deep Inference impulsionou o desenvolvimento de ferramentas para implementação de Redes Neurais Convolucionais (CNNs, do inglês Convolutional Neural Networks) como bibliotecas específicas para o desenvolvimento de CNNs, a popularização também estimulou o desenvolvimento de GPUs com recursos de aceleração de CNNs. Os desenvolvedores e cientistas que trabalham com CNNs precisam de estudos científicos experimentais que apontem qual a biblioteca mais adequada para determinada microarquitetura de GPU. Objetivos: Comparar o desempenho das bibliotecas CNTK, PyTorch, TensorFlow 1.15 e TensorFlow 2.2 em diferentes microarquiteturas de GPU (Kepler, Maxwell, Pascal e Turing) utilizando a CNN LeNet-5 e o dataset MNIST. Metodologia: Inicialmente, foi realizado um mapeamento sistemático como forma de identificar e sistematizar os principais benchmarks, em seguida foi realizado um experimento avaliando o impacto da API Keras no desempenho das bibliotecas utilizadas no estudo comparativo e por último foram realizados quatro experimentos comparando as bibliotecas, cada experimento foi realizado em um ambiente acelerado por uma GPU com uma microarquitetura diferente. Resultados: Para o mapeamento sistemático, foi identificado que as arquiteturas clássicas de CNN como LeNet-5 e AlexNet são as mais utilizadas como benchmarks, também mostrou que os datasets mais utilizados em benchmarking são o ImageNet, MNIST e CIFAR-10. O experimento sobre o impacto da API Keras mostrou que a API impacta negativamente no desempenho de todas as bibliotecas testadas. No estudo comparativo de desempenho das bibliotecas, a biblioteca PyTorch apresentou o pior desempenho e as bibliotecas CNTK, TensorFlow 1.15 e TensorFlow 2.2 alternaram entre os três menores tempos de execução. Conclusão: Os resultados evidenciam que a biblioteca PyTorch apresenta um baixo nível de utilização da GPU e utiliza uma grande quantidade de memória se comparada as outras bibliotecas que apresentaram um desempenho superior. São Cristóvão, SE more...

This deliverable reports on the progress of the SODALITE project in developing a model for application and infrastructure performance. Application performance deployed in an infrastructure can be predicted from the Performance Model of the application and infrastructure. Predicting and improving performance helps not only to reduce application costs but also improve infrastructure utilisation in general. As part of this M12 deliverable, this report describes the prototype model developed and also reports on the baseline performance of use case applications. more...

This project shall help to bring a tool to fight against bullying in schools. It is also possible to use it in different scenes where a camera is recording a common area shared by people, such as companies, banks, prisons, or hospitals. To achieve that, the issue is approached from two main modules. The first one, a comparative study of approaches to detect violence in video, using image and video analyser Neural Networks (NN)s: a custom image analyser NN based on LeNet5, AlexNet, custom stacked long short-term memory (LSTM) and convolutional LSTM based NNs. The trainings are done with two datasets that have been subject to modifications to correct possible misinterpretations during the learning and pretraining is applied. The LeNet5 based NN is unsuccessful and tested with an independent dataset AlexNet is inaccurate. The best results are obtained with a stacked LSTM NN and a convolutional LSTM with dropout and a LSTM layer. Both NNs achieve over 90 % of accuracy with training and validation datasets, meanwhile the stacked LSTM and the convolutional NN achieve, respectively, 75 % and 100 % of accuracy with a small independent test dataset created. The convolutional LSTM needed 10 times less epochs to achieve the same result as the stacked LSTM. The second module consists of a violence detection system that applies the best solution obtained from the comparative study. The violence detection system saves the frames detected as violence with date, time and camera name and emits a sound alarm when more than a certain number of consecutive frames are evaluated as containing violence. This way the sensitivity of the system is reduced and avoids false alarms due to small mistakes done by the intelligence. more...

Virtuelle Desktops mittels Grafikkarten auf der Serverseite zu beschleunigen, ist längst keine Nische für besonders leistungshungrige Spezialanwendungen mehr. Auch typische Office-Nutzer profitieren signifikant davon, wenn unter der Haube ihrer Desktops im Rechenzentrum zusätzlich zur CPU auch Grafikkarten ihren Dienst tun. Der Artikel erläutert die Grundlagen der GPU-Beschleunigung für virtuelle Desktops. more...

The aim of this chapter is to present the automated fingerprint recognition technology and its use for forensic applications. After a brief historical review, we provide an introduction to modern Automated Fingerprint Identification Systems (AFIS) by discussing their functionalities and accuracy. The topic then becomes more technical and goes through some of the recently introduced approaches for fingerprint recognition (both for fingerprint and fingermarks). Forensic applications exploiting the recognition of fingerprints (identity verification and identification) and fingermarks (forensic intelligence, investigation and evaluation) are then described. Finally, a discussion about the current topics and foreseeable challenges in terms of technology and application concludes the chapter. more...

Privacy amplification is an indispensable step in postprocessing of continuous-variable quantum key distribution (CV-QKD), which is used to distill unconditional secure keys from identical corrected keys between two distant legal parties. The processing speed of privacy amplification has a significant effect on the secret key rate of CV-QKD system. We report the high-speed parallel implementation of length-compatible privacy amplification algorithm based on graphic processing unit. Length-compatible algorithm is used to satisfy the security requirement of privacy amplification at different transmission distances when considering finite-size effect. We achieve the speed of privacy amplification over 1 Gbps at arbitrary input length and the speed is one to two orders of magnitude faster than previous demonstrations, which supports high-speed real-time CV-QKD system and ensures the security of privacy amplification., 10 pages, 3 figures more...

[ES] El Deep Learning (o aprendizaje profundo) es un término que está tomando mucha fuerza en muchas de las ramas de la ciencia a día de hoy. La facilidad de generar modelos de Inteligencia Artificial con entornos de Deep Learning como TensorFlow ha hecho que la divulgación de este tópico se este extendiendo de manera notable. En esta memoria, pretendemos introducir el concepto de redes neuronales, de Deep Learning, su historia y los factores por los que está teniendo tanto impacto en la actualidad. En concreto, revisaremos la necesidad del hardware para la producción de estos modelos y qué tipos de entornos están a nuestra disposición para realizar modelos de redes neuronales. A continuación introduciremos TensorFlow, la librería de Deep Learning desarrollada por Google, y la utilizaremos para formular nuestros modelos que nos servirán para resolver los problemas de reconocimiento de imágenes que se proponen en las bases de datos (o datasets) de MNIST y CIFAR-10. Finalmente, ejecutaremos y entrenaremos nuestros modelos con las GPUs proporcionadas por la Universidad. Comentaremos y compararemos resultados con ambos datasets, y presentaremos nuestras conclusiones del trabajo realizado y de los resultados conseguidos., [CA] El Deep Learning (o Aprenentatge Profund) es un terme que està prenent molta força en moltes de les branques de la ciència hui en día. La facilitat de generar models de Inteligència Artificial amb entorns de Deep Learning com TensorFlow ha fet que la divulgació d’aquest tòpic s’estiga escampant per tot arreu. En aquesta memòria pretenem introduir el concepte de xarxes neuronals y el Deep Learning, la seua història y els factors pels que hui en día ha tingut èxit. En concret, revisarem la necesitat de hardware per a la producció d’aquests models y quins tipus d’entorns estàn al nostre abast per a realitzar models de xarxes neuronals. Acte seguit introduirem TensorFlow, la llibrería de Deep Learning desenvolupada per Google, y la utilitzarem per a formular models que ens serviràn per a resoldre els problemes de reconeixement d’imatges que ens proposen els datasets de MNIST y CIFAR-10. Finalment, executarem y entrenarem els nostres models amb unes GPUs proveïdes per l’Universitat. Comentarem i compararem resultats amb ambdós bases de dades (o datasets), y presentarem les nostres conclusions del treball realitzat y del resultats aconseguits., [EN] The concept of Deep Learning is growing in almost every branch of science, and for a good reason. The ease with which someone can produce Artificial Intelligence models with the resources and frameworks available nowadays is making the term Deep Learning echo almost everywhere with an astonishing speed. In this report, we aim to introduce the concept of a Neural Network and bind it with Deep Learning and Image Classification, learn its own history and explain the reason why it is getting so much importance nowadays. To be precise, we will review the need of hardware to develop such models and what kind of environments are available for us to produce Neural Network Models. Later we will introduce TensorFlow, the Deep Learning framework developed by Google. We will use TensorFlow to code our models in order to solve the Image Classification problems provided by the datasets of MNIST and CIFAR10. Lastly, we will execute and train our models with the GPUs provided by our University. Later, we will discuss/compare the results from both datasets, and then finally explain the conclusions from our work and the data obtained. more...

The goal of this research report is to compare the program execution times using the Central Processing Unit and NVIDIA’s high throughput solution CUDA to execute the program on the Graphic Processing Unit. The programs will run an algorithm to convert a raster image into dotted half-tone. CUDA`s programming model allows programmers to run their programs on a parallel GPU architecture consisting of several blocks of threads that are being ran in parallel. This dynamic technology allows to further decrease the time it takes to run a program, by simply adding more GPU’s or upgrading the existing one The research experiment concluded that NVIDIA solution is faster than using usual CPU serial code more...

Due the recent increase of the volume of data that has been generated, organizing this data has become one of the biggest problems in Computer Science. Among the different strategies propose to deal efficiently and effectively for this purpose, we highlight those related to clustering, more specifically, density-based clustering strategies, which stands out for its ability to define clusters of arbitrary shape and the robustness to deal with the presence of data noise, such as DBSCAN and OPTICS. However, these algorithms are still a computational challenge since they are distance-based proposals. In this work we present a new approach to make OPTICS feasible based on data indexing strategy. Although the simplicity with which the data are indexed, using graphs, it allows explore various parallelization opportunities, which were explored using graphic processing unit (GPU). Based on this structure, the complexity of OPTICS is reduced to O(E *logV ) in the worst case, becoming itself very fast. In our evaluation we show that our proposal can be over 200x faster than its sequential version using CPU. more...

Blade is one of the core components of turbine machinery. The reliability of blade is directly related to the normal operation of plant unit. However, with the increase of blade length and flow rate, non-linear effects such as finite deformation must be considered in strength computation to guarantee enough accuracy. Parallel computation is adopted to improve the efficiency of classical nonlinear finite element method and shorten the blade design period. So it is of extraordinary importance for engineering practice. In this paper, the dynamic partial differential equations and the finite element method forms for turbine blades under centrifugal load and flow load are given firstly. Then, according to the characteristics of turbine blade model, the classical method is optimized based on central processing unit + graphics processing unit heterogeneous parallel computation. Finally, the numerical experiment validations are performed. The computation speed of the algorithm proposed in this paper is compared with the speed of ANSYS. For the rectangle plate model with mesh number of 10 k to 4000 k, a maximum speed-up of 4.31 can be obtained. For the real blade-rim model with mesh number of 500 k, the speed-up of 4.54 times can be obtained. more...

Orientador: Prof. Dr. Thales Souza Dissertação (mestrado) - Universidade Federal do ABC, Programa de Pós-Graduação em Engenharia Elétrica, Santo André, 2018. Este trabalho apresenta um estudo de estabilidade estática de tensão que utiliza o fluxo de carga continuado para determinação das curvas PV. A determinação das curvas PV permite a obtenção da margem de estabilidade de tensão que consiste na determinação da quantidade de carga adicional, a partir de um ponto de operação, que causaria o colapso de tensão no sistema. Ademais, com o objetivo de verificar a influência na rede elétrica, o estudo proposto considera a inserção de centrais geradoras eólicas como representação de energia renovável e intermitente. Nesse sentido, de maneira a buscar uma otimização computacional e numérica do algoritmo utilizado para o processo de simulação, é empregada a computação paralela como uma ferramenta de otimização computacional de forma que possam ser executadas em diferentes unidades de processamento ao mesmo tempo, reduzindo o tempo total de execução. Para tanto, é apresentada uma metodologia de programação paralela em Unidades de Processamento Gráfico (GPU), utilizando a plataforma de programação CUDA, aplicada ao fluxo de carga continuado nos cálculos computacionais intensivos e a CPU, encarregada de comandar a sequência do algoritmo e executar cálculos menores. Assim, a partir das simulações propostas, é realizada uma comparação de desempenho e precisão entre utilizar ou não o processamento paralelo, com fim de otimizar os estudos de estabilidade requeridos no planejamento e operação do sistema elétrico. This work is focused in a study of static voltage stability that uses the Continuation Power Flow (CPF) to determine the PV curves. The determination of the PV curves allows obtaining the Voltage Stability Margin, which consists in determining the amount of additional load from an operation point that would cause the voltage collapse in the system. In addition, the proposed study considers the insertion of Wind Energy Conversion Systems (WECS) as a representation of renewable energy, with the objective of verifying its influence in the electric network. In doing so, to achieve a computational and numerical optimization of the algorithm used for the simulation process, parallel computing is used as a computational optimization tool, so that it can be executed in different processing units simultaneously, reducing the total execution time. In this way, a methodology of parallel programming in Graphics Processing Units (GPU) is presented using the platform of programming CUDA, applied to the Continuation Power Flow in the intensive computational calculations and the CPU, in charge for commanding the sequence of the algorithm and performing smaller calculations. Thus, from the proposed simulations, a comparison of performance and precision parallel processing and non-parallel processing is performed, in order to optimize the stability studies required in electrical system planning and operation. more...

This paper proposes a hybrid gate-level leakage model for the use with the Monte Carlo (MC) analysis approach, which combines a lookup table (LUT) model with a first-order exponential-polynomial model (first-order model, herein). For the process parameters having strong nonlinear relationships with the logarithm of leakage current, the proposed model uses the LUT approach for the sake of modeling accuracy. For the other process parameters, it uses the first-order model for increased efficiency. During the library characterization for each type of logic gates, the proposed approach determines the process parameters for which it will use the LUT model. And, it determines the number of LUT data points, which can maximize analysis efficiency with acceptable accuracy, based on the user-defined threshold. The proposed model was implemented for gate-level MC leakage analysis using three graphic processing units. In experiments, the proposed approach exhibited the average errors of ¡5% in both mean and standard deviation with reference to SPICE-level MC leakage analysis. In comparison, MC analysis with the first-order model exhibited more than 90% errors. In CPU times, the proposed hybrid approach took only two to five times longer runtimes. In comparison with the full LUT model, the proposed hybrid model was up to one hundred times faster while increasing the average errors by only 3%. Finally, the proposed approach completed a leakage analysis of an OpenSparc T2 core of 4.5 million gates with a runtime of . more...

In CBCT, image reconstruction is difficult to meet the requirements of the user real-time reconstruction because data amount of image reconstruction is large, operation complexity is high and the time of reconstruction is long. CUDA based on GPU launched by NVIDIA is very suitable for large-scale parallel computing problem. This paper optimizes cone beam CT reconstruction algorithm by CUDA and improves the speed of weighted back-projection and filtering, and shortens the data access time by using the texture memory and constant memory in CUDA to respectively store the kernel function and the filtered data. The experimental results show that the reconstruction speed and the reconstruction quality are obviously improved compared with the reconstruction method based on GPU. more...

In the drug discovery field, solving the problem of virtual screening is a long term-goal. The scoring functionality which evaluates the fitness of the docking result is one of the major challenges in virtual screening. In general, scoring functionality in docking requires large amount of floating-point calculations and usually takes several weeks or even months to be finished. This time-consuming disadvantage is unacceptable especially when highly fatal and infectious virus arises such as SARS and H1N1. This paper presents how to leverage the computational power of GPU to accelerate Dock6 [1]’s Amber [2] scoring with NVIDIA CUDA [3] platform. We also discuss many factors that will greatly influence the performance after porting the Amber scoring to GPU, including thread management, data transfer and divergence hidden. Our GPU implementation shows a 6.5x speedup with respect to the original version running on AMD dual-core CPU for the same problem size. more...

Višeskalni model mišićne kontrakcije obuhvata fizičke i biohemijske procese koji se odvijaju na više prostornih i vremenskih skala. Zbog velike računske složenosti, ovakvi modeli mogu biti upotrebljivi samo primenom tehnika paralelizacije i upotrebom računarskih resursa visokih performansi. Postojeća rešenja za ubrzavanje višeskalnih simulacija mišića su oslonjena na univerzalne softverske okvire za paralelizaciju višeskalnih modela. Ona podrazumevaju upotrebu skupe računarske opreme visokih performansi, što stvara visoku cenu troškova eksploatacije i čini preskupom njihovu upotrebu u svakodnevnoj kliničkoj praksi. Pored toga, nijedno od rešenja ne koristi nikakvu metodologiju kojom bi se specifična znanja o domenu iskoristila u svrhu boljeg iskorišćenja raspoloživih računarskih resursa. Ova disertacija je rezultat istraživanja vezanih za prevazilaženje ova dva nedostatka postojećih rešenja. U disertaciji je uveden koncept distribuiranja proračuna višeskalnih modela mišića u okruženju sastavljenom od standardnih procesorskih jezgara i grafičkih procesorskih jedinica. Koncept je primenjen u kontekstu dvoskalnog modela mišića, u kojem je model konačnih elemenata na makroskali uparen sa Haksli modelom kinetike poprečnih mostova na mikro skali. Paralelno rešenje je bazirano na statičkoj dekompoziciji domena mikromodela. Uvedeni koncept je implementiran kao softverska platforma, Mexie, koja predstavlja prenosivo i proširivo softversko rešenje modularne arhitekture. Platforma koristi specifičan metod raspoređivanja poslova prilagođen heterogenom izvršnom okruženju, gde se u obzir uzimaju različite brzine izračunavanja i memorijska ograničenja resursa. Algoritam raspodele uzima u obzir i različitost u računskoj kompleksnosti mikromodela. Kako kompleksnost mikromodela ne može biti precizno određena pre izvršenja dvoskalne simulacije, definisana je metodologija za procenu. Metodologija je bazirana na upotrebi Hilovog fenomenološkog modela za predviđanje stanja i ulaznih parametara mikromodela tokom dvoskalne simulacije. Na osnovu procenjenih vrednosti, specifičan alat za procenu, obučen tehnikama mašinskog učenja, procenjuje računsku kompleksnost svih mikromodela. Implementirano rešenje je verifikovano na testnim modelima, kao i na realnom modelu, pokazujući visoku iskorišćenost raspoloživih resursa, čvrsto skaliranje i ubrzanja od dva reda veličine u odnosu na sekvencijalnu varijantu. Zahvaljujući efikasnosti modela i značajnom poboljšanju odnosa cena/performanse, otvoren je put za nova istraživanja u oblasti fiziologije mišića i napravljen značajan korak u smeru upotrebe višeskalnih modela mišića u budućoj kliničkoj praksi. Multi-scale models of muscle contraction rely on the integration of physical and biochemical properties across multiple length and time scales, consequentially they are highly CPU consuming. Feasible usage of these models can be reached only by employing high performance computing environments. Existing efforts for accelerating multi-scale muscle simulations lean on general purpose parallelization techniques and frameworks. These solutions imply the usage of expensive large-scale computational resources, which produces overwhelming costs for the everyday practical application of such models. Additionally, none of them employ any methodology based on specific model domain knowledge to provide customized resource utilization. The thesis resulted from the investigations addressed to both of these deficiencies. The dissertation introduces concept of the distributed calculations of the multi-scale muscle models in a mixed CPU–GPU environment in order to improve computational speed within a reasonable budget. The concept is applied to a two-scale muscle model, in which a finite element macro model is coupled with the microscopic Huxley kinetics model. The parallel solution is based on decomposition of the micro model domain and static scheduling policy. Computations related to Huxley models run on both CPUs and GPUs, while finite element calculations are executed sequentially and solely on the CPU. Introduced concept is implemented as a software platform, Mexie, which represents a portable and extensible software solution with modular architecture. The platform uses a specific load balancing method customized for heterogeneous execution environments. It takes into account differences in computational speeds and memory limits of available computing resources. Additionally, the dispatch algorithm considers computational weights of each micro model, providing more balanced workload. Since these computational weights could not be known before exact two-scale simulation run, a novel methodology for assessing their values is introduced. The methodology uses simple Hill phenomenological model in order to predict model states and input parameters of all micro models during two-scale simulation. These estimated data is then used by specific predictor tool, trained by machine learning techniques, for assessing micro model computational complexities. Implemented solution was verified on both benchmark and real-world example, showing high utilization of involved processing units, strong scalability and speed-up of two orders of magnitude compared to the sequential CPU run. This major improvement in computational feasibility and significant lowering of price-to-performance ratio of the multi-scale muscle models paves the way for new discoveries in the field of muscle modeling and usage of shuch models in the future clinical applications, as well. more...

The problem of microelectromechanical systems (MEMS) development for near-wall turbulent flow control has been considered. Complex solution based on MEMS and graphic processing platform NVIDIA Jetson TK1 has been proposed that demonstrates the development of powerful embedded application to process the stream of data from the streamlined surface. It will help to reach better maneuverability, increase the range of aircraft payload capability by reducing the turbulent drag due to control system that includes combination of on-chip information modules, sensors and actuators. more...

Mobile Cloud Computing (MCC) combines mobile computing and cloud computing aiming to aid performance of mobile devices. The idea is simple: thin devices offload heavy methods to resource-rich servers in the clouds. We believe that in the near future MCC will adopt more advanced offloading techniques. In particular, in this paper we envision a scenario where offloading frameworks will have to deal with GPU code offloading. Amazon already offers instances with Graphics Processing Units (GPU), which can be used for this purpose. We propose and implement MCC-Adviser, a simulation tool that can predict the performance of GPUs with different number of cores using Stochastic Petri Nets. We tested MCC-Adviser in a case study with one of the expensive Amazon GPU instances. The simulations showed that it is possible to minimize costs, while satisfying user's quality of service requirements, by utilizing less powerful instances. more...

Desde años recientes, el paralelismo hace parte de la arquitectura de las computadoras personales al incluir unidades de co-procesamiento como las unidades de procesamiento gráfico, para conformar así una arquitectura heterogénea. Este artículo presenta la implementación de algoritmos de enjambres sobre esta arquitectura para resolver problemas de optimización de funciones, destacando su estructura inherentemente paralela y sus propiedades de control distribuido. En estos algoritmos se paralelizan los individuos de la población y las dimensiones del problema gracias a la granuralidad del sistema de procesamiento, que además proporciona una baja latencia de comunicaciones entre los individuos debido al procesamiento embebido. Para evaluar las potencialidades de los algoritmos de enjambres sobre la plataforma heterogénea, son implementados dos de ellos: el algoritmo de enjambre de partículas y el algoritmo de enjambre de bacterias. Se utiliza la aceleración como métrica para contrastar los algoritmos en la arquitectura heterogénea compuesta por una GPU NVIDIA GTX480 y una unidad de procesamiento secuencial, donde el algoritmo de enjambre de partículas obtiene una aceleración de hasta 36,82x y el algoritmo de enjambre de bacterias logra una aceleración de hasta 9,26x. Además, se evalúa el efecto al incrementar el tamaño en las poblaciones donde la aceleración es significativamente diferenciable pero con riesgos en la calidad de las soluciones. Since few years ago, the parallel processing has been embedded in personal computers by including co-processing units as the graphics processing units resulting in a heterogeneous platform. This paper presents the implementation of swarm algorithms on this platform to solve several functions from optimization problems, where they highlight their inherent parallel processing and distributed control features. In the swarm algorithms, each individual and dimension problem are parallelized by the granularity of the processing system which also offer low communication latency between individuals through the embedded processing. To evaluate the potential of swarm algorithms on graphics processing units we have implemented two of them: the particle swarm optimization algorithm and the bacterial foraging optimization algorithm. The algorithms’ performance is measured using the acceleration where they are contrasted between a typical sequential processing platform and the NVIDIA GeForce GTX480 heterogeneous platform; the results show that the particle swarm algorithm obtained up to 36.82x and the bacterial foraging swarm algorithm obtained up to 9.26x. Finally, the effect to increase the size of the population is evaluated where we show both the dispersion and the quality of the solutions are decreased despite of high acceleration performance since the initial distribution of the individuals can converge to local optimal solution. more...

The availability of cost-effective data collections and storage hardware has allowed organizations to accumulate very large data sets, which are a potential source of previously unknown valuable information. The process of discovering interesting patterns in such large data sets is referred to as data mining. Outlier detection is a data mining task consisting in the discovery of observations which deviate substantially from the rest of the data, and has many important practical applications. Outlier detection in very large data sets is however computationally very demanding and currently requires highperformance computing facilities. We propose a family of parallel algorithms for Graphic Processing Units (GPU), derived from two distance-based outlier detection algorithms: the BruteForce and the SolvingSet. We analyze their performance with an extensive set of experiments, comparing the GPU implementations with the base CPU versions and obtaining significant speedups. more...

International audience; This paper is devoted to pricing American options using Monte Carlo and Malliavin calculus. We develop this method on two types of models, the multidimensional exponential model with deterministic (nonconstant) volatility and the multidimensional Heston model. To obtain good numerical results, we introduce a variance reduction technique based on conditioning and a bias reduction method that relies on an appropriate choice of the number of simulated paths in the computation of the quotient of two expectations. Since our techniques are well suited to parallel implementation, our numerical experiments are performed using multicore central process unit and many-core graphic processing unit environments. more...

Date of Conference: 4-6 May 2009 Conference Name: 3DTV Conference: The True Vision - Capture, Transmission and Display of 3D Video, IEEE 2009 A holographic fringe pattern generation methods is based on Fraunhofer diffraction and subsequent segmentation and approximation of the fringe pattern. Several modifications of the original algorithm are already proposed to improve the quality of reconstructions. We compare the quality of to the reconstructed images from different versions of this algorithm by taking the reconstructions from the Fresnel hologram as a reference. Since, there is not any generally accepted objective quality assessment method for such reconstructions, we used some experimental methods such as intensity spread over the reconstructed images, total noise power, and peak-signal-to-noise for comparison. Then we chose the best performing algorithm in terms of ireconstruction quality, and developed a GPU-based implementation to accelerate the computation speed. The quality of the resultant reconstructions is comparable to reconstructions from Fresnel holograms; much higher speed is achieved due to multi-GPU implemetation. more...

Membrane computing is a theoretical model of computation that inspired from the structure and functioning of cells. Membrane computing models naturally have parallel structure. Most of the simulations of membrane computing have been done in a serial way on a machine with a central processing unit (CPU). This has neglected the advantage of parallelism in membrane computing. This paper uses Graphic Processing Unit (GPU) as a parallel tool to implement membrane computing. The method minimizes data transferring which is time consuming procedure between the device and the host by processing all computing on GPU and transfer only the final results to CPU. Simulations show that speed increases up to 15 times compared to sequential simulation and using of shared memory increases speed, up to 38 times. more...

These latest years witnessed an impressive improvement of graphics hardware both in terms of features and in terms of computational power. This improvement can be easily observed in computer games, where effects which, until few years ago, could only be achieved with expensive CPU computation are now shown interactively. Although the GPU has been designed for implementing graphics effects, it is still it basically a processing unit with its own memory, and, being specialized for algebraic tasks, supplies a number of floating point operations per second which is orders of magnitude greater than the CPU. This suggested to the graphics community that the GPU could also be used for general purpose computation and a number of papers have been published on how to hack the GPU to this target. Following this trend we propose a framework for using GPU for implementing techniques for deformable objects represented as generic meshes. The framework only assumes than the global computation is the union of local computations, which is true for all the explicit methods. more...

Deep learning algorithms base their success on building high learning capacity models with millions of parameters that are tuned in a data-driven fashion. These models are trained by processing millions of examples, so that the development of more accurate algorithms is usually limited by the throughput of the computing devices on which they are trained. This project show how the training of a state-of-the-art neural network for computer vision can be parallelized on a distributed GPU cluster, Minotauro GPU cluster from Barcelona Supercomputing Center with the TensorFlow framework. In this project, two approaches for distributed training are used, the synchronous and the mixed-asynchronous. The effect of distributing the training process is addressed from two different points of view. First, the scalability of the task and its performance in the distributed setting are analyzed. Second, the impact of distributed training methods on the final accuracy of the models is studied. The results show an improvement for both focused areas. On one hand, the experiments show promising results in order to train a neural network faster. The training time is decreased from 106 hours to 16 hours in mixedasynchronous and 12 hours in synchronous. On the other hand we can observe how increasing the numbers of GPUs in one node rises the throughput, images per second, in a near-linear way. Moreover the accuracy can be maintained, like the one node training, in the synchronous methods. more...