http://sedici.unlp.edu.ar:80/handle/10915/152373 Tue, 03 Nov 2026 01:32:20 GMT 2026-11-03T01:32:20Z http://sedici.unlp.edu.ar:80/handle/10915/153546 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) La paralelización de algoritmos con fuerte dependencia de datos no logra grandes mejoras de rendindento eo un entorno de memoria distribuida debido a las sincronizaciones, donde gran parte de los nodos guedan a la espera de datos procesados por otro/s nodo/s. Los nodos en espera pueden realizar cómputos parciales mientras esperan los datos que generan la dependencia siempre que dispongan de datos para poder realizarios y que se preserven los resultados parciales hasta que sean necesarios. Se presenta los primeros resultados de aplicar esta técnica de paralelización sobre el algoritmo de factorización de Cholesky con mejora de hasta un tercio en el tiempo de ejecución. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/153546 2010-01-01T00:00:00Z La paralelización de algoritmos con fuerte dependencia de datos no logra grandes mejoras de rendindento eo un entorno de memoria distribuida debido a las sincronizaciones, donde gran parte de los nodos guedan a la espera de datos procesados por otro/s nodo/s. Los nodos en espera pueden realizar cómputos parciales mientras esperan los datos que generan la dependencia siempre que dispongan de datos para poder realizarios y que se preserven los resultados parciales hasta que sean necesarios. Se presenta los primeros resultados de aplicar esta técnica de paralelización sobre el algoritmo de factorización de Cholesky con mejora de hasta un tercio en el tiempo de ejecución. http://sedici.unlp.edu.ar:80/handle/10915/152739 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) UltraSPARC T2 is a system on a chip. It contains 8 cores and it is capable of running 64 simultaneous hardware threads. Although it was designed for server workloads with high throughput requirements, we tested the architecture using a visual computing application. In this work we perform image binarization in video streams. We use a serial and a parallel implementation to test the performance of the parallelization. An evaluation of the speedup for several screen resolutions of the same video stream shows that this architecture can also be a good choice for real-time video streaming and processing applications. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152739 2010-01-01T00:00:00Z UltraSPARC T2 is a system on a chip. It contains 8 cores and it is capable of running 64 simultaneous hardware threads. Although it was designed for server workloads with high throughput requirements, we tested the architecture using a visual computing application. In this work we perform image binarization in video streams. We use a serial and a parallel implementation to test the performance of the parallelization. An evaluation of the speedup for several screen resolutions of the same video stream shows that this architecture can also be a good choice for real-time video streaming and processing applications. http://sedici.unlp.edu.ar:80/handle/10915/152738 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) In the last few years, several millions of high-capable mobile devices were sold around the globe. Therefore, integrating mobile devices to Grid systems may increase significantly their resources. However, mobile devices have more limited resources than personal computers or servers. In particular, minimizing energy consumption is very important because mobile devices rely on battery as energy supply. We propose a novel job scheduler that aims to use the energy in an efficient way. To implement it, we developed a simple but effective battery estimation model. In some experiments our energy aware scheduler outperformed traditional Grid schedulers to effectively address energy constrains in mobile Grids. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152738 2010-01-01T00:00:00Z In the last few years, several millions of high-capable mobile devices were sold around the globe. Therefore, integrating mobile devices to Grid systems may increase significantly their resources. However, mobile devices have more limited resources than personal computers or servers. In particular, minimizing energy consumption is very important because mobile devices rely on battery as energy supply. We propose a novel job scheduler that aims to use the energy in an efficient way. To implement it, we developed a simple but effective battery estimation model. In some experiments our energy aware scheduler outperformed traditional Grid schedulers to effectively address energy constrains in mobile Grids. http://sedici.unlp.edu.ar:80/handle/10915/152736 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) Multicore processors have opened new paths for improving the parallel performance in cluster environments. Nevertheless, the selection of different combinations between the amount of nodes and the number of cores per node implies different results in terms of parallel performance. We performed an impact assessment on the parallel performance of node-core combinations using a parallel approach of a machine learning ensemble algorithm. Our results reveal that two key factors for selecting a suitable node-core combination: the network capabilities and the workload distribution. We observed that the network interconnection limits the amount of nodes that can be efficiently used, due to the extranode communications does not allow to keep scaling as the number of nodes is increased. The best results were obtained by reaching a balance between intra-node and extra-node communications. By the other hand, the parallel performance can be negatively affected when the workload distribution is not homogeneous among nodes. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152736 2010-01-01T00:00:00Z Multicore processors have opened new paths for improving the parallel performance in cluster environments. Nevertheless, the selection of different combinations between the amount of nodes and the number of cores per node implies different results in terms of parallel performance. We performed an impact assessment on the parallel performance of node-core combinations using a parallel approach of a machine learning ensemble algorithm. Our results reveal that two key factors for selecting a suitable node-core combination: the network capabilities and the workload distribution. We observed that the network interconnection limits the amount of nodes that can be efficiently used, due to the extranode communications does not allow to keep scaling as the number of nodes is increased. The best results were obtained by reaching a balance between intra-node and extra-node communications. By the other hand, the parallel performance can be negatively affected when the workload distribution is not homogeneous among nodes. http://sedici.unlp.edu.ar:80/handle/10915/152735 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) As new HPC technologies appear, small local research laboratories face uncertain options when building hybrid clusters. Users may find difficult to choose among several multicore products with different core densities. Our research project intends to build knowledge about HPC problems to be able to help local researchers. We analyze NUMA hardware for use in clusters and present a case study. We run a well-known benchmark over MPI and advise the user depending on application features. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152735 2010-01-01T00:00:00Z As new HPC technologies appear, small local research laboratories face uncertain options when building hybrid clusters. Users may find difficult to choose among several multicore products with different core densities. Our research project intends to build knowledge about HPC problems to be able to help local researchers. We analyze NUMA hardware for use in clusters and present a case study. We run a well-known benchmark over MPI and advise the user depending on application features. http://sedici.unlp.edu.ar:80/handle/10915/152733 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) This paper presents a workflow in order to generate and collect execution information in multi-process environments, which are suitable for simulating and studying private and shared cache organizations. We have developed a main tool that makes dynamic instrumentation with PIN, a controller to administrate the execution of processes and three workflow application modules to generate programs traces, simulate cache memories and manage the execution of several programs in parallel. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152733 2010-01-01T00:00:00Z This paper presents a workflow in order to generate and collect execution information in multi-process environments, which are suitable for simulating and studying private and shared cache organizations. We have developed a main tool that makes dynamic instrumentation with PIN, a controller to administrate the execution of processes and three workflow application modules to generate programs traces, simulate cache memories and manage the execution of several programs in parallel. http://sedici.unlp.edu.ar:80/handle/10915/152732 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) Condor is a middleware specially design to manage job queues in dedicated and non-dedicated infrastructure. Condor seems to be a suitable tool to be used when the lack of available dedicated resources represent an impediment to run large application problems. This paper discusses the use of Condor in processing MPI applications problems in homogeneous and heterogeneous pools of nodes. In particular several experiments with different computing and network usage have been performed. From these tests it is concluded that Condor introduces a negligible overhead and also allows to conform scalable pools of nodes with a good cost-benefit ratio. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152732 2010-01-01T00:00:00Z Condor is a middleware specially design to manage job queues in dedicated and non-dedicated infrastructure. Condor seems to be a suitable tool to be used when the lack of available dedicated resources represent an impediment to run large application problems. This paper discusses the use of Condor in processing MPI applications problems in homogeneous and heterogeneous pools of nodes. In particular several experiments with different computing and network usage have been performed. From these tests it is concluded that Condor introduces a negligible overhead and also allows to conform scalable pools of nodes with a good cost-benefit ratio. http://sedici.unlp.edu.ar:80/handle/10915/152731 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) Solving the structure of protein-protein complexes is one of the most important tasks in structural biology. Even though there has been great progress in recent years there still a small number of protein complexes structures deposited in the Protein Data Bank in comparison to isolated partners. In this sense, the computational prediction of protein complexes starting from the unbound structures, protein-protein Docking algorithms, has emerged as a reasonable alternative. Many docking programs employ Fast Fourier Transform (FFT) correlations as an efficient search strategy. We describe an implementation of a protein-protein docking program based on FFT surface complementarity that runs entirely on a Graphics Processing Unit (GPU), including grid generation, rotation and scoring. We evaluate its performance, and show that it can be up to 13 times faster than conventional CPU based implementations. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152731 2010-01-01T00:00:00Z Solving the structure of protein-protein complexes is one of the most important tasks in structural biology. Even though there has been great progress in recent years there still a small number of protein complexes structures deposited in the Protein Data Bank in comparison to isolated partners. In this sense, the computational prediction of protein complexes starting from the unbound structures, protein-protein Docking algorithms, has emerged as a reasonable alternative. Many docking programs employ Fast Fourier Transform (FFT) correlations as an efficient search strategy. We describe an implementation of a protein-protein docking program based on FFT surface complementarity that runs entirely on a Graphics Processing Unit (GPU), including grid generation, rotation and scoring. We evaluate its performance, and show that it can be up to 13 times faster than conventional CPU based implementations. http://sedici.unlp.edu.ar:80/handle/10915/152730 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) The Multilevel algorithm (ML) has been applied successfully as a metaheuristic for different combinatorial optimization problems: Graph Partitioning, Traveling Salesman, Graph Coloring, see refs. [6,7,18]. The main difficulty of ML are the convergence times needed to obtain solutions at a distance of 7% - 5% to the best known solution in large scale problems. In order to reduce these convergence times we studied numerically a Parallel Multilevel heuristic with Neural Network partitioning and uncoarsening + refinement phases (PML+PNN) for the Graph Bisection Problem on geometrically connected graphs. Our main result establish that for graphs with n∊[4000,12000] vertices, the performance of the parallel ML+NN heuristic increases linearly as n increases with respect to the parallel ML heuristic. For n∊{10000,12000} the distance to the best solution found is 0.32,0.25 respectively that is obtained with a quadratic computing time. This suggests improving the performance of the PML+PNN heuristic by means of a hill climbing improvement heuristic. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152730 2010-01-01T00:00:00Z The Multilevel algorithm (ML) has been applied successfully as a metaheuristic for different combinatorial optimization problems: Graph Partitioning, Traveling Salesman, Graph Coloring, see refs. [6,7,18]. The main difficulty of ML are the convergence times needed to obtain solutions at a distance of 7% - 5% to the best known solution in large scale problems. In order to reduce these convergence times we studied numerically a Parallel Multilevel heuristic with Neural Network partitioning and uncoarsening + refinement phases (PML+PNN) for the Graph Bisection Problem on geometrically connected graphs. Our main result establish that for graphs with n∊[4000,12000] vertices, the performance of the parallel ML+NN heuristic increases linearly as n increases with respect to the parallel ML heuristic. For n∊{10000,12000} the distance to the best solution found is 0.32,0.25 respectively that is obtained with a quadratic computing time. This suggests improving the performance of the PML+PNN heuristic by means of a hill climbing improvement heuristic. http://sedici.unlp.edu.ar:80/handle/10915/152728 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) Graphics Processing Units (GPUs) is an emergent hardware technology potentially suitable for High Performance Applications. In this work GPU processing of Molecular Dynamics (MD) simulations is discussed. An overview of GPU architecture is provided and its main features are outlined. Some relevant atomistic MD applications are processed using GPU hardware. The same applications are simulated using CPU oriented applications as well in order to obtain comparable results. In this case both serial and MPI codes are considered. Wall-clock timing of GPU results obtained with a relatively low-end card are found to be comparable to timing from multiple core clusters. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152728 2010-01-01T00:00:00Z Graphics Processing Units (GPUs) is an emergent hardware technology potentially suitable for High Performance Applications. In this work GPU processing of Molecular Dynamics (MD) simulations is discussed. An overview of GPU architecture is provided and its main features are outlined. Some relevant atomistic MD applications are processed using GPU hardware. The same applications are simulated using CPU oriented applications as well in order to obtain comparable results. In this case both serial and MPI codes are considered. Wall-clock timing of GPU results obtained with a relatively low-end card are found to be comparable to timing from multiple core clusters. http://sedici.unlp.edu.ar:80/handle/10915/152727 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) In several scientific and business environments, the magnitude of problems require the execution of complex applications, like work-flows. Condor includes DAGMan extension in order to manage workflows. Such kind of applications are very popular because jobs can be reused and distributed. Scheduling is a central issue in the assignation of resources to workflows. It can be seen as the process of assigning the resources to the workflow's jobs in a convenient way. This work presents a performance-driven approach to make these assignments (also known as mappings). The mappings generated determine in which resources the jobs must be executed. The execution stage is delegated to the DAGMan Workflow System. This approach is compared with the standalone execution of the workflow over DAGMan. Experimental results show that the performance-driven approach improves the execution time of workflows in comparison with the standalone DAGMan approach. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152727 2010-01-01T00:00:00Z In several scientific and business environments, the magnitude of problems require the execution of complex applications, like work-flows. Condor includes DAGMan extension in order to manage workflows. Such kind of applications are very popular because jobs can be reused and distributed. Scheduling is a central issue in the assignation of resources to workflows. It can be seen as the process of assigning the resources to the workflow's jobs in a convenient way. This work presents a performance-driven approach to make these assignments (also known as mappings). The mappings generated determine in which resources the jobs must be executed. The execution stage is delegated to the DAGMan Workflow System. This approach is compared with the standalone execution of the workflow over DAGMan. Experimental results show that the performance-driven approach improves the execution time of workflows in comparison with the standalone DAGMan approach. http://sedici.unlp.edu.ar:80/handle/10915/152638 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) The proliferation of huge computer centers with thousands of processors available, summed up with the new computing paradigms as cloud computing, make distributed environment a maze of new perspectives. Dynamics of resource information is increasing in complexity, and nowadays is almost impossible to be captured using a static hierarchy. Consequently, we focused on developing new tools to attack this new challenges. Grid-Matrix is a framework with an user-friendly GUI, based on the famous grid simulator SimGrid. It is focused on the analysis of information propagation policies, including automatic generation of very varied network topologies, and many different policies scenarios. As a case study, we present the generation of clique and ring networks under three different policies: hierarchical, super-peer and random. These preliminary results shows the potential of this powerful framework, plenty of features for the profound analysis of information propagation. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152638 2010-01-01T00:00:00Z The proliferation of huge computer centers with thousands of processors available, summed up with the new computing paradigms as cloud computing, make distributed environment a maze of new perspectives. Dynamics of resource information is increasing in complexity, and nowadays is almost impossible to be captured using a static hierarchy. Consequently, we focused on developing new tools to attack this new challenges. Grid-Matrix is a framework with an user-friendly GUI, based on the famous grid simulator SimGrid. It is focused on the analysis of information propagation policies, including automatic generation of very varied network topologies, and many different policies scenarios. As a case study, we present the generation of clique and ring networks under three different policies: hierarchical, super-peer and random. These preliminary results shows the potential of this powerful framework, plenty of features for the profound analysis of information propagation. http://sedici.unlp.edu.ar:80/handle/10915/152637 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) We study two different techniques for the computation of a matrix inverse, the traditional approach based on Gaussian factorization and the Gauss-Jordan elimination alternative more suitable for parallel architectures. The target architecture is a current general-purpose multi-core processor (CPU) connected to a graphics processor (GPU). Parallelism is obtained from the use of libraries MKL (for the CPU) and CUBLAS (for the GPU), as well as, performing simultaneously operations in both architectures. Numerical experiments performed on a system equipped with two Intel QuadCore processors and a Tesla C1060 GPU, illustrate the efficiency attained by the Gauss-Jordan elimination implementation. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152637 2010-01-01T00:00:00Z We study two different techniques for the computation of a matrix inverse, the traditional approach based on Gaussian factorization and the Gauss-Jordan elimination alternative more suitable for parallel architectures. The target architecture is a current general-purpose multi-core processor (CPU) connected to a graphics processor (GPU). Parallelism is obtained from the use of libraries MKL (for the CPU) and CUBLAS (for the GPU), as well as, performing simultaneously operations in both architectures. Numerical experiments performed on a system equipped with two Intel QuadCore processors and a Tesla C1060 GPU, illustrate the efficiency attained by the Gauss-Jordan elimination implementation. http://sedici.unlp.edu.ar:80/handle/10915/152634 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) In this work, we present the design and implementation of a highly modular and flexible software framework to implement numerical models based on the finite element method (FEM) and its extension to deal with distributed problems. This work improves the current implementation by the adition of parallel calculations capabilities by means of the substructure technique applied to solve problems by the FEM in clusters of computers using the MPI protocol. We considered the solution of a general Poisson problem as a test case to conduct experiments in order to evaluate the scaling capabilities of our code. Conclusions are extracted with focus on future lines of development. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152634 2010-01-01T00:00:00Z In this work, we present the design and implementation of a highly modular and flexible software framework to implement numerical models based on the finite element method (FEM) and its extension to deal with distributed problems. This work improves the current implementation by the adition of parallel calculations capabilities by means of the substructure technique applied to solve problems by the FEM in clusters of computers using the MPI protocol. We considered the solution of a general Poisson problem as a test case to conduct experiments in order to evaluate the scaling capabilities of our code. Conclusions are extracted with focus on future lines of development. http://sedici.unlp.edu.ar:80/handle/10915/152633 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) The k-th nearest neighbour problem for 3D data has been widely studied, nevertheless, the surge of using GPU (Graphical Processing Unit) as general-purpose computing units opens up the need to design and implement new algorithms, that allow us to get results more rapidly than using conventional algorithms. The main advantage of GPU is its great computing capacity for parallel computing. This is due to an architecture that contemplates a great number of processing cores originally designed to graphics processing and that can currently be used for other purposes such as high performance cientific calculation. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152633 2010-01-01T00:00:00Z The k-th nearest neighbour problem for 3D data has been widely studied, nevertheless, the surge of using GPU (Graphical Processing Unit) as general-purpose computing units opens up the need to design and implement new algorithms, that allow us to get results more rapidly than using conventional algorithms. The main advantage of GPU is its great computing capacity for parallel computing. This is due to an architecture that contemplates a great number of processing cores originally designed to graphics processing and that can currently be used for other purposes such as high performance cientific calculation. http://sedici.unlp.edu.ar:80/handle/10915/152631 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) The proliferation of chess servers on the Internet has turned active chess, blitz and lightning, into a vast cognitive phenomenon involving engaged participants. Here we use this large database of human decision making (rapid chess) as a privileged window to understand human cognition. FICS (Free Internet Chess Server), http://www.freechess.org/ is a free ICS-compatible server for playing chess games through Internet, with more than 300.000 registered users. Using this available chess server in the Internet, we constructed a massive decision-making database. This data includes thousands of million moves of chess games, with the estimated time of each one of them. In order to evaluate the goodness of moves, we used Crafty (an open-source chess engine) to analyse the score of the move. This process is compute expensive, so we parallelized the analysis on a Beowulf cluster. We studied the structure of the time players take to make a move during a game, and using parallelization we were able to analyse a huge amount of moves obtaining a quantification of the quality of the decision made in millions of instances. This approach allowed us to identify a number of statistical fingerprints that uniquely characterize the emergent structure of the game. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152631 2010-01-01T00:00:00Z The proliferation of chess servers on the Internet has turned active chess, blitz and lightning, into a vast cognitive phenomenon involving engaged participants. Here we use this large database of human decision making (rapid chess) as a privileged window to understand human cognition. FICS (Free Internet Chess Server), http://www.freechess.org/ is a free ICS-compatible server for playing chess games through Internet, with more than 300.000 registered users. Using this available chess server in the Internet, we constructed a massive decision-making database. This data includes thousands of million moves of chess games, with the estimated time of each one of them. In order to evaluate the goodness of moves, we used Crafty (an open-source chess engine) to analyse the score of the move. This process is compute expensive, so we parallelized the analysis on a Beowulf cluster. We studied the structure of the time players take to make a move during a game, and using parallelization we were able to analyse a huge amount of moves obtaining a quantification of the quality of the decision made in millions of instances. This approach allowed us to identify a number of statistical fingerprints that uniquely characterize the emergent structure of the game. http://sedici.unlp.edu.ar:80/handle/10915/152630 Objeto de conferencia High-Performance Computing Symposium (HPC 2010) - JAIIO 39 (UADE, 30 de agosto al 3 de septiembre de 2010) This article presents the application of parallel computing techniques using Graphic Processing Unit (GPU) in order to improve the computational efficiency of numerical methods applied to uid dynamics problems. In the last ten years, GPUs have emerged as a major paradigm for solving complex problems using parallel computing techniques. Fluid dynamics problems usually requires large execution times to perform simulations for realistic scenarios. In this work, two numerical models for fluid dynamics are presented, and parallel implementations on GPU for the Strongly Implicit Procedure and the Cyclic Reduction methods for solving linear systems are introduced. The experimental evaluation of the proposed methods demonstrates that a significant reduction on the computing times can be attained when solving linear systems with representative dimensions, and preliminary results show that the efficiency gains also propagate to the numerical models for fluid dynamics. Fri, 01 Jan 2010 00:00:00 GMT http://sedici.unlp.edu.ar:80/handle/10915/152630 2010-01-01T00:00:00Z This article presents the application of parallel computing techniques using Graphic Processing Unit (GPU) in order to improve the computational efficiency of numerical methods applied to uid dynamics problems. In the last ten years, GPUs have emerged as a major paradigm for solving complex problems using parallel computing techniques. Fluid dynamics problems usually requires large execution times to perform simulations for realistic scenarios. In this work, two numerical models for fluid dynamics are presented, and parallel implementations on GPU for the Strongly Implicit Procedure and the Cyclic Reduction methods for solving linear systems are introduced. The experimental evaluation of the proposed methods demonstrates that a significant reduction on the computing times can be attained when solving linear systems with representative dimensions, and preliminary results show that the efficiency gains also propagate to the numerical models for fluid dynamics.