Search results

The scientific community in its never-ending road of larger and more efficient computational resources is in need of more efficient implementations that can adapt efficiently on the current parallel platforms. Graphics processing units are an appropriate platform that cover some of these demands. This architecture presents a high performance with a reduced cost and an efficient power consumption. However, the memory capacity in these devices is reduced and so expensive memory transfers are necessary to deal with big problems. Today, the lattice-Boltzmann method (LBM) has positioned as an efficient approach for Computational Fluid Dynamics simulations. Despite this method is particularly amenable to be efficiently parallelized, it is in need of a considerable memory capacity, which is the consequence of a dramatic fall in performance when dealing with large simulations. In this work, we propose some initiatives to minimize such demand of memory, which allows us to execute bigger simulations on the same platform without additional memory transfers, keeping a high performance. In particular, we present 2 new implementations, LBM-Ghost and LBM-Swap, which are deeply analyzed, presenting the pros and cons of each of them. ; This project was funded by the Spanish Ministry of Economy and Competitiveness (MINECO): BCAM Severo Ochoa accreditation SEV-2013-0323, MTM2013-40824, Computación de Altas Prestaciones VII TIN2015-65316-P, by the Basque Excellence Research Center (BERC 2014-2017) pro- gram by the Basque Government, and by the Departament d' Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programació i Entorns d' Execució Paral·lels (2014-SGR-1051). We also thank the support of the computing facilities of Extremadura Research Centre for Advanced Technologies (CETA-CIEMAT) and NVIDIA GPU Research Center program for the provided resources, as well as the support of NVIDIA through the BSC/UPC NVIDIA GPU Center of Excellence. ; Peer Reviewed ; Postprint (author's final draft)

The scientific community in its never-ending road of larger and more efficient computational resources is in need of more efficient implementations that can adapt efficiently on the current parallel platforms. Graphics processing units are an appropriate platform that cover some of these demands. This architecture presents a high performance with a reduced cost and an efficient power consumption. However, the memory capacity in these devices is reduced and so expensive memory transfers are necessary to deal with big problems. Today, the lattice-Boltzmann method (LBM) has positioned as an efficient approach for Computational Fluid Dynamics simulations. Despite this method is particularly amenable to be efficiently parallelized, it is in need of a considerable memory capacity, which is the consequence of a dramatic fall in performance when dealing with large simulations. In this work, we propose some initiatives to minimize such demand of memory, which allows us to execute bigger simulations on the same platform without additional memory transfers, keeping a high performance. In particular, we present 2 new implementations, LBM-Ghost and LBM-Swap, which are deeply analyzed, presenting the pros and cons of each of them. ; This project was funded by the Spanish Ministry of Economy and Competitiveness (MINECO): BCAM Severo Ochoa accreditation SEV-2013-0323, MTM2013-40824, Computación de Altas Prestaciones VII TIN2015-65316-P, by the Basque Excellence Research Center (BERC 2014-2017) pro- gram by the Basque Government, and by the Departament d' Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programació i Entorns d' Execució Paral·lels (2014-SGR-1051). We also thank the support of the computing facilities of Extremadura Research Centre for Advanced Technologies (CETA-CIEMAT) and NVIDIA GPU Research Center program for the provided resources, as well as the support of NVIDIA through the BSC/UPC NVIDIA GPU Center of Excellence. ; Peer Reviewed ; Postprint (author's final draft)

Today one of the most important challenges in HPC is the development of computers with a low power consumption. In this context, recently, new embedded many-core systems have emerged. One of them is Kalray. Unlike other many-core architectures, Kalray is not a co-processor (self-hosted). One interesting feature of the Kalray architecture is the Network on Chip (NoC) connection. Habitually, the communication in many-core architectures is carried out via shared memory. However, in Kalray, the communication among processing elements can also be via Message-Passing on the NoC. One of the main motivations of this work is to present the main constraints to deal with the Kalray architecture. In particular, we focused on memory management and communication. We assess the use of NoC and shared memory on Kalray. Unlike shared memory, the implementation of Message-Passing on NoC is not transparent from programmer point of view. The synchronization among processing elements and NoC is other of the challenges to deal with in the Karlay processor. Although the synchronization using Message-Passing is more complex and consuming time than using shared memory, we obtain an overall speedup close to 6 when using Message-Passing on NoC with respect to the use of shared memory. Additionally, we have measured the power consumption of both approaches. Despite of being faster, the use of NoC presents a higher power consumption with respect to the approach that exploits shared memory. This additional consumption in Watts is about a 50%. However, the reduction in time by using NoC has an important impact on the overall power consumption as well. ; This research has been supported by EU-FET grant EUNISON 308874, the Basque Excellence Research Center (BERC 2014-2017) program by the Basque Government, the projects founded by the Spanish Ministry of Economy and Competitiveness (MINECO): BCAM Severo Ochoa accreditation SEV-2013-0323, MTM2013-40824 and TIN2015-65316-P. We acknowledge Research Center in Real-TIme & Embedded Computing Systes - CISTER for the provided resources. ; Postprint (published version)

Today one of the most important challenges in HPC is the development of computers with a low power consumption. In this context, recently, new embedded many-core systems have emerged. One of them is Kalray. Unlike other many-core architectures, Kalray is not a co-processor (self-hosted). One interesting feature of the Kalray architecture is the Network on Chip (NoC) connection. Habitually, the communication in many-core architectures is carried out via shared memory. However, in Kalray, the communication among processing elements can also be via Message-Passing on the NoC. One of the main motivations of this work is to present the main constraints to deal with the Kalray architecture. In particular, we focused on memory management and communication. We assess the use of NoC and shared memory on Kalray. Unlike shared memory, the implementation of Message-Passing on NoC is not transparent from programmer point of view. The synchronization among processing elements and NoC is other of the challenges to deal with in the Karlay processor. Although the synchronization using Message-Passing is more complex and consuming time than using shared memory, we obtain an overall speedup close to 6 when using Message-Passing on NoC with respect to the use of shared memory. Additionally, we have measured the power consumption of both approaches. Despite of being faster, the use of NoC presents a higher power consumption with respect to the approach that exploits shared memory. This additional consumption in Watts is about a 50%. However, the reduction in time by using NoC has an important impact on the overall power consumption as well. ; This research has been supported by EU-FET grant EUNISON 308874, the Basque Excellence Research Center (BERC 2014-2017) program by the Basque Government, the projects founded by the Spanish Ministry of Economy and Competitiveness (MINECO): BCAM Severo Ochoa accreditation SEV-2013-0323, MTM2013-40824 and TIN2015-65316-P. We acknowledge Research Center in Real-TIme & Embedded Computing Systes - CISTER for the provided resources. ; Postprint (published version)

Convolutional neural networks (CNNs) have recently attracted considerable attention due to their outstanding accuracy in applications, such as image recognition and natural language processing. While one advantage of the CNNs over other types of neural networks is their reduced computational cost, faster execution is still desired for both training and inference. Since convolution operations pose most of the execution time, multiple algorithms were and are being developed with the aim of accelerating this type of operations. However, due to the wide range of convolution parameter configurations used in the CNNs and the possible data type representations, it is not straightforward to assess in advance which of the available algorithms will be the best performing in each particular case. In this paper, we present a performance evaluation of the convolution algorithms provided by the cuDNN, the library used by most deep learning frameworks for their GPU operations. In our analysis, we leverage the convolution parameter configurations from widely used the CNNs and discuss which algorithms are better suited depending on the convolution parameters for both 32 and 16-bit floating-point (FP) data representations. Our results show that the filter size and the number of inputs are the most significant parameters when selecting a GPU convolution algorithm for 32-bit FP data. For 16-bit FP, leveraging specialized arithmetic units (NVIDIA Tensor Cores) is key to obtain the best performance. ; This work was supported by the European Union's Horizon 2020 Research and Innovation Program under the Marie Sklodowska-Curie under Grant 749516, and in part by the Spanish Juan de la Cierva under Grant IJCI-2017-33511 ; Peer Reviewed ; Postprint (published version)

The main contribution of this work is to increase the coding productivity for GPU programming by using the concept of Static Graphs. To do so, we have combined the new CUDA Graph API with the OpenACC programming model. We use as test cases a well-known and widely used problems in HPC and AI: the Particle Swarm Optimization. We complement the OpenACC functionality with the use of CUDA Graph, achieving accelerations of more than one order of magnitude, and a performance very close to a reference and optimized CUDA code. Finally, we propose a new specification to incorporate the concept of Static Graphs into the OpenACC specification. ; This project has received funding from the EPEEC project from the European Union's Horizon 2020 Research and Innovation program under grant agreement No. 801051. ; Peer Reviewed ; Postprint (author's final draft)

Convolutional neural networks (CNNs) have recently attracted considerable attention due to their outstanding accuracy in applications, such as image recognition and natural language processing. While one advantage of the CNNs over other types of neural networks is their reduced computational cost, faster execution is still desired for both training and inference. Since convolution operations pose most of the execution time, multiple algorithms were and are being developed with the aim of accelerating this type of operations. However, due to the wide range of convolution parameter configurations used in the CNNs and the possible data type representations, it is not straightforward to assess in advance which of the available algorithms will be the best performing in each particular case. In this paper, we present a performance evaluation of the convolution algorithms provided by the cuDNN, the library used by most deep learning frameworks for their GPU operations. In our analysis, we leverage the convolution parameter configurations from widely used the CNNs and discuss which algorithms are better suited depending on the convolution parameters for both 32 and 16-bit floating-point (FP) data representations. Our results show that the filter size and the number of inputs are the most significant parameters when selecting a GPU convolution algorithm for 32-bit FP data. For 16-bit FP, leveraging specialized arithmetic units (NVIDIA Tensor Cores) is key to obtain the best performance. ; This work was supported by the European Union's Horizon 2020 Research and Innovation Program under the Marie Sklodowska-Curie under Grant 749516, and in part by the Spanish Juan de la Cierva under Grant IJCI-2017-33511 ; Peer Reviewed ; Postprint (published version)

The main contribution of this work is to increase the coding productivity of GPU programming by using the concept of Static Graphs. GPU capabilities have been increasing significantly in terms of performance and memory capacity. However, there are still some problems in terms of scalability and limitations to the amount of work that a GPU can perform at a time. To minimize the overhead associated with the launch of GPU kernels, as well as to maximize the use of GPU capacity, we have combined the new CUDA Graph API with the CUDA programming model (including CUDA math libraries) and the OpenACC programming model. We use as test cases two different, well-known and widely used problems in HPC and AI: the Conjugate Gradient method and the Particle Swarm Optimization. In the first test case (Conjugate Gradient) we focus on the integration of Static Graphs with CUDA. In this case, we are able to significantly outperform the NVIDIA reference code, reaching an acceleration of up to 11× thanks to a better implementation, which can benefit from the new CUDA Graph capabilities. In the second test case (Particle Swarm Optimization), we complement the OpenACC functionality with the use of CUDA Graph, achieving again accelerations of up to one order of magnitude, with average speedups ranging from 2× to 4×, and performance very close to a reference and optimized CUDA code. Our main target is to achieve a higher coding productivity model for GPU programming by using Static Graphs, which provides, in a very transparent way, a better exploitation of the GPU capacity. The combination of using Static Graphs with two of the current most important GPU programming models (CUDA and OpenACC) is able to reduce considerably the execution time w.r.t. the use of CUDA and OpenACC only, achieving accelerations of up to more than one order of magnitude. Finally, we propose an interface to incorporate the concept of Static Graphs into the OpenACC Specifications. ; his research was funded by EPEEC project from the European Union's Horizon 2020 Research and Innovation program under grant agreement No. 801051. This manuscript has been authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The US government retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan, accessed on 13 April 2022). ; Peer Reviewed ; Postprint (published version)

The simulation of the behavior of the human brain is one of the most ambitious challenges today with a non-end of important applications. We can find many different initiatives in the USA, Europe and Japan which attempt to achieve such a challenging target. In this work, we focus on the most important European initiative (the Human Brain Project) and on one of the models developed in this project. This tool simulates the spikes triggered in a neural network by computing the voltage capacitance on the neurons' morphology, being one of the most precise simulators today. In the present work, we have evaluated the use of MPI+OpenMP tasking on top of this framework. We prove that this approach is able to achieve a good scaling even when computing a relatively low workload (number of neurons) per node. One of our targets consists of achieving not only a highly scalable implementation, but also to develop a tool with a high degree of abstraction without losing control and performance by using MPI+OpenMP tasking. The main motivation of this work is the evaluation of this cutting-edge simulation on multi-morphology neural networks. The simulation of a high number of neurons, which are completely different among them, is an important challenge. In fact, in the multi-morphology simulations, we find an important unbalancing between the nodes, mainly due to the differences in the neurons, which causes an important under-utilization of the available resources. In this work, the authors present and evaluate mechanisms to deal with this and reduce the time of this kind of simulations considerably. ; We would like to appreciate the valuable feedback and help provided by the main developers of the Arbor simulator: Benjamin Cumming (ETH Zürich) and Alexander Peyser (Jülich Supercomputing Center). This project has also received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 720270 (HBP SGA1 and SGA2), from the Spanish Ministry of Economy and Competitiveness under the project Computación de Altas Prestaciones VII (TIN2015-65316- P) and the Departament d'Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programació i Entorns d'Execució Paral · lels (2014-SGR-1051). Finally, this project also received funding from the Spanish Ministry of Economy and Competitiveness under the Juan de la Cierva Grant Agreement No IJCI-2017-33511, and from the European Union's Horizon 2020 research and innovation program under the Marie Sklodowska Curie grant agreement No. 749516. ; Peer Reviewed ; Postprint (author's final draft) ; Postprint (author's final draft)

The solving of tridiagonal systems is one of the most computationally expensive parts in many applications, so that multiple studies have explored the use of NVIDIA GPUs to accelerate such computation. However, these studies have mainly focused on using parallel algorithms to compute such systems, which can efficiently exploit the shared memory and are able to saturate the GPUs capacity with a low number of systems, presenting a poor scalability when dealing with a relatively high number of systems. The gtsvStridedBatch routine in the cuSPARSE NVIDIA package is one of these examples, which is used as reference in this article. We propose a new implementation (cuThomasBatch) based on the Thomas algorithm. Unlike other algorithms, the Thomas algorithm is sequential, and so a coarse-grained approach is implemented where one CUDA thread solves a complete tridiagonal system instead of one CUDA block as in gtsvStridedBatch. To achieve a good scalability using this approach, it is necessary to carry out a transformation in the way that the inputs are stored in memory to exploit coalescence (contiguous threads access to contiguous memory locations). Different variants regarding the transformation of the data are explored in detail. We also explore some variants for the case of variable batch, when the size of the systems of the batch has different size (cuThomasVBatch). The results given in this study prove that the implementations carried out in this work are able to beat the reference code, being up to 5× (in double precision) and 6× (in single precision) faster using the latest NVIDIA GPU architecture, the Pascal P100. ; This project was funded from the European Union's Horizon 2020 research and innovation programme under grant agreement 720270 (HBPSGA1), from the Spanish Ministry of Economy and Competitiveness under the project Computación de Altas Prestaciones VII (TIN2015-65316-P)and the Departament d'Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programació iEntorns d'Execució Paral·lels (2014-SGR-1051). We thank the support of NVIDIA through the BSC/UPC NVIDIA GPU Center of Excellence andthe valuable feedback provided by Lung Sheng Chien and Alex Fit-Florea. Antonio J. Peña was cofinanced by the Spanish Ministry of Economy andCompetitiveness under Juan de la Cierva fellowship number IJCI-2015-23266. ; Peer Reviewed ; Postprint (author's final draft)

A current trend in high-performance computing is to decompose a large linear algebra problem into batches containing thousands of smaller problems, that can be solved independently, before collating the results. To standardize the interface to these routines, the community is developing an extension to the BLAS standard (the batched BLAS), enabling users to perform thousands of small BLAS operations in parallel whilst making efficient use of their hardware. We discuss the benefits and drawbacks of the current batched BLAS proposals and perform a number of experiments, focusing on a general matrix-matrix multiplication (GEMM), to explore their affect on the performance. In particular we analyze the effect of novel data layouts which, for example, interleave the matrices in memory to aid vectorization and prefetching of data. Utilizing these modifications our code outperforms both MKL1 CuBLAS2 by up to 6 times on the self-hosted Intel KNL (codenamed Knights Landing) and Kepler GPU architectures, for large numbers of double precision GEMM operations using matrices of size 2 × 2 to 20 × 20. ; The authors would like to thank The University of Tennessee for the use of their computational resources. This research was funded in part from the European Union's Horizon 2020 research and innovation programme under the NLAFET grant agreement No. 671633. ; Peer Reviewed ; Postprint (published version)

A current trend in high-performance computing is to decompose a large linear algebra problem into batches containing thousands of smaller problems, that can be solved independently, before collating the results. To standardize the interface to these routines, the community is developing an extension to the BLAS standard (the batched BLAS), enabling users to perform thousands of small BLAS operations in parallel whilst making efficient use of their hardware. We discuss the benefits and drawbacks of the current batched BLAS proposals and perform a number of experiments, focusing on a general matrix-matrix multiplication (GEMM), to explore their affect on the performance. In particular we analyze the effect of novel data layouts which, for example, interleave the matrices in memory to aid vectorization and prefetching of data. Utilizing these modifications our code outperforms both MKL1 CuBLAS2 by up to 6 times on the self-hosted Intel KNL (codenamed Knights Landing) and Kepler GPU architectures, for large numbers of double precision GEMM operations using matrices of size 2 × 2 to 20 × 20. ; The authors would like to thank The University of Tennessee for the use of their computational resources. This research was funded in part from the European Union's Horizon 2020 research and innovation programme under the NLAFET grant agreement No. 671633. ; Peer Reviewed ; Postprint (published version)

The simulation of the behavior of the Human Brain is one of the most important challenges in computing today. The main problem consists of finding efficient ways to manipulate and compute the huge volume of data that this kind of simulations need, using the current technology. In this sense, this work is focused on one of the main steps of such simulation, which consists of computing the Voltage on neurons' morphology. This is carried out using the Hines Algorithm and, although this algorithm is the optimum method in terms of number of operations, it is in need of non-trivial modifications to be efficiently parallelized on GPUs. We proposed several optimizations to accelerate this algorithm on GPU-based architectures, exploring the limitations of both, method and architecture, to be able to solve efficiently a high number of Hines systems (neurons). Each of the optimizations are deeply analyzed and described. Two different approaches are studied, one for mono-morphology simulations (batch of neurons with the same shape) and one for multi-morphology simulations (batch of neurons where every neuron has a different shape). In mono-morphology simulations we obtain a good performance using just a single kernel to compute all the neurons. However this turns out to be inefficient on multi-morphology simulations. Unlike the previous scenario, in multi-morphology simulations a much more complex implementation is necessary to obtain a good performance. In this case, we must execute more than one single GPU kernel. In every execution (kernel call) one specific part of the batch of the neurons is solved. These parts can be seen as multiple and independent tridiagonal systems. Although the present paper is focused on the simulation of the behavior of the Human Brain, some of these techniques, in particular those related to the solving of tridiagonal systems, can be also used for multiple oil and gas simulations. Our studies have proven that the optimizations proposed in the present work can achieve high performance on those computations with a high number of neurons, being our GPU implementations about 4× and 8× faster than the OpenMP multicore implementation (16 cores), using one and two NVIDIA K80 GPUs respectively. Also, it is important to highlight that these optimizations can continue scaling, even when dealing with a very high number of neurons. ; This project has received funding from the European Union's Horizon 2020 Research and Innovation Programme under Grant Agreement No. 720270 (HBP SGA1), from the Spanish Ministry of Economy and Competitiveness under the project Computación de Altas Prestaciones VII (TIN2015-65316-P), the Departament d'Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programació i Entorns d'Execució Parallels (2014-SGR-1051). We thank the support of NVIDIA through the BSC/UPC NVIDIA GPU Center of Excellence, and the European Union's Horizon 2020 Research and Innovation Program under the Marie Sklodowska-Curie Grant Agreement No. 749516. ; Peer Reviewed ; Postprint (published version)

The simulation of the behavior of the Human Brain is one of the most important challenges in computing today. The main problem consists of finding efficient ways to manipulate and compute the huge volume of data that this kind of simulations need, using the current technology. In this sense, this work is focused on one of the main steps of such simulation, which consists of computing the Voltage on neurons' morphology. This is carried out using the Hines Algorithm and, although this algorithm is the optimum method in terms of number of operations, it is in need of non-trivial modifications to be efficiently parallelized on GPUs. We proposed several optimizations to accelerate this algorithm on GPU-based architectures, exploring the limitations of both, method and architecture, to be able to solve efficiently a high number of Hines systems (neurons). Each of the optimizations are deeply analyzed and described. Two different approaches are studied, one for mono-morphology simulations (batch of neurons with the same shape) and one for multi-morphology simulations (batch of neurons where every neuron has a different shape). In mono-morphology simulations we obtain a good performance using just a single kernel to compute all the neurons. However this turns out to be inefficient on multi-morphology simulations. Unlike the previous scenario, in multi-morphology simulations a much more complex implementation is necessary to obtain a good performance. In this case, we must execute more than one single GPU kernel. In every execution (kernel call) one specific part of the batch of the neurons is solved. These parts can be seen as multiple and independent tridiagonal systems. Although the present paper is focused on the simulation of the behavior of the Human Brain, some of these techniques, in particular those related to the solving of tridiagonal systems, can be also used for multiple oil and gas simulations. Our studies have proven that the optimizations proposed in the present work can achieve high performance on those computations with a high number of neurons, being our GPU implementations about 4× and 8× faster than the OpenMP multicore implementation (16 cores), using one and two NVIDIA K80 GPUs respectively. Also, it is important to highlight that these optimizations can continue scaling, even when dealing with a very high number of neurons. ; This project has received funding from the European Union's Horizon 2020 Research and Innovation Programme under Grant Agreement No. 720270 (HBP SGA1), from the Spanish Ministry of Economy and Competitiveness under the project Computación de Altas Prestaciones VII (TIN2015-65316-P), the Departament d'Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programació i Entorns d'Execució Parallels (2014-SGR-1051). We thank the support of NVIDIA through the BSC/UPC NVIDIA GPU Center of Excellence, and the European Union's Horizon 2020 Research and Innovation Program under the Marie Sklodowska-Curie Grant Agreement No. 749516. ; Peer Reviewed ; Postprint (published version)

The simulation of the behavior of the Human Brain is one of the most important challenges today in computing. The main problem consists of finding efficient ways to manipulate and compute the huge volume of data that this kind of simulations need, using the current technology. In this sense, this work is focused on one of the main steps of such simulation, which consists of computing the Voltage on neurons' morphology. This is carried out using the Hines Algorithm. Although this algorithm is the optimum method in terms of number of operations, it is in need of non-trivial modifications to be efficiently parallelized on NVIDIA GPUs. We proposed several optimizations to accelerate this algorithm on GPU-based architectures, exploring the limitations of both, method and architecture, to be able to solve efficiently a high number of Hines systems (neurons). Each of the optimizations are deeply analyzed and described. To evaluate the impact of the optimizations on real inputs, we have used 6 different morphologies in terms of size and branches. Our studies have proven that the optimizations proposed in the present work can achieve a high performance on those computations with a high number of neurons, being our GPU implementations about 4× and 8× faster than the OpenMP multicore implementation (16 cores), using one and two K80 NVIDIA GPUs respectively. Also, it is important to highlight that these optimizations can continue scaling even when dealing with number of neurons. ; This project has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 720270 (HBP SGA1), from the Spanish Ministry of Economy and Competitiveness under the project Computación de Altas Prestaciones VII (TIN2015-65316-P) and the Departament d'Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programació i Entorns d'Execució Paral·lels (2014-SGR-1051). We thank the support of NVIDIA through the BSC/UPC NVIDIA GPU Center of Excellence. Antonio J. Peña is cofinanced by the Spanish Ministry of Economy and Competitiveness under Juan de la Cierva fellowship number IJCI-2015-23266. ; Peer Reviewed ; Postprint (published version)