Search NASA⌕ Search

SEARCH · Search NASA

Results for “Load Balancing”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 127 records · Page 7
All sourcesNASA NTRSDOE OSTI
DOE OSTI2015

Particle Communication and Domain Neighbor Coupling: Scalable Domain Decomposed Algorithms for Monte Carlo Particle Transport

In order to run Monte Carlo particle transport calculations on new supercomputers with hundreds of thousands or millions of processors, care must be taken to implement scalable algorithms. This means that the algorithms must continue to perform well as the processor count increases. In this paper, we examine the scalability of:(1) globally resolving the particle locations on the correct processor, (2) deciding that particle streaming communication has finished, and (3) efficiently coupling neighbor domains together with different replication levels. We have run domain decomposed Monte Carlo particle transport on up to 2 21 = 2,097,152 MPI processes on the IBM BG/Q Sequoia supercomputer and observed scalable results that agree with our theoretical predictions. These calculations were carefully constructed to have the same amount of work on every processor, i.e. the calculation is already load balanced. We also examine load imbalanced calculations where each domain’s replication level is proportional to its particle workload. In this case we show how to efficiently couple together adjacent domains to maintain within workgroup load balance and minimize memory usage.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2021

Asynchronous distributed-memory task-parallel algorithm for compressible flows on unstructured 3D Eulerian grids

Here, we discuss the implementation of a finite element method, used to numerically solve the Euler equations of compressible flows, using an asynchronous runtime system (RTS). The algorithm is implemented for distributed-memory machines, using stationary unstructured 3D meshes, combining data-, and task-parallelism on top of the Charm++ RTS. Charm++’s execution model is asynchronous by default, allowing arbitrary overlap of computation and communication. Task-parallelism allows scheduling parts of an algorithm independently of, or dependent on, each other. Built-in automatic load balancing enables continuous redistribution of computational load by migration of work units based on real-time CPU load measurement. The RTS also features automatic checkpointing, fault tolerance, resilience against hardware failure, and supports power-, and energy-aware computation. We demonstrate scalability up to 25 x 10 9 cells at $\mathscr{O}$10 4 compute cores and the benefits of automatic load balancing for irregular workloads. The full source code with documentation is available at https://quinoacomputing.org.

42 ENGINEERING↗
DOE OSTI2022

Evaluation of Optimal Net Load Management in Microgrids Using Hardware-in-the-Loop Simulation: Preprint

This paper presents the performance evaluation of a net load management (NLM) engine that balances load and generation in an isolated community to power a critical facility after a grid interruption event (e.g., lost a large generation unit). The NLM engine is a central dispatch control system that provides high-speed cost-optimal coordination of a set of net loads (combination of generation and deferrable loads) connected to the same electrical network. The NLM algorithm cost-optimally dispatches the grid-following inverters and deferrable loads based on the demanded power and load priorities, and the grid-forming inverters use droop control to form system voltages and share active and reactive power. A controller-hardware-in-the-loop platform is developed to evaluate the control performance of the NLM algorithm. The experimental results indicate that the NLM engine can achieve the targeted system voltage and frequency and balance load and generation to serve the critical facility.

droop control↗
DOE OSTI2022

Evaluation of Optimal Net Load Management in Microgrids Using Hardware-in-the-Loop Simulation

This paper presents the performance evaluation of a net load management (NLM) engine that balances load and generation in an isolated community to power a critical facility after a grid interruption event (e.g., the loss of a large generation unit). This NLM engine is particularly important for microgrid systems because it provides a high-speed, cost-optimal control solution to coordinate grid-forming inverters and to dispatch grid-following inverters and deferrable loads in microgrid systems to enhance grid resilience and reliability. The NLM algorithm cost-optimally dispatches the grid-following inverters and deferrable loads based on the demanded power and load priorities, and the grid-forming inverters use droop control to form system voltages and share active and reactive power. A controller-hardware-in-the-loop platform is developed to evaluate the control performance of the NLM algorithm with two sequential contingency events of lost generation units. The experimental results indicate that the NLM engine can maintain system stability, achieve the targeted system voltage and frequency, and balance load and generation to serve the critical facility with improved system resilience and reliability.

grid-following inverter↗
DOE OSTI2022

Evaluation of Optimal Net Load Management in Microgrids Using Hardware-in-the-Loop Simulation

This presentation discusses the performance evaluation of a net load management (NLM) engine that balances load and generation in an isolated community to power a critical facility after a grid interruption event (e.g., the loss of a large generation unit). This NLM engine is particularly important for microgrid systems because it provides a high-speed, cost-optimal control solution to coordinate grid-forming inverters and to dispatch grid-following inverters and deferrable loads in microgrid systems to enhance grid resilience and reliability. The NLM algorithm cost-optimally dispatches the grid-following inverters and deferrable loads based on the demanded power and load priorities, and the grid-forming inverters use droop control to form system voltages and share active and reactive power. A controller-hardware-in-the-loop platform is developed to evaluate the control performance of the NLM algorithm with two sequential contingency events of lost generation units. The experimental results indicate that the NLM engine can maintain system stability, achieve the targeted system voltage and frequency, and balance load and generation to serve the critical facility with improved system resilience and reliability.

droop control↗
DOE OSTI2016

Work stealing for GPU-accelerated parallel programs in a global address space framework

Task parallelism is an attractive approach to automatically load balance the computation in a parallel system and adapt to dynamism exhibited by parallel systems. Exploiting task parallelism through work stealing has been extensively studied in shared and distributed-memory contexts. In this paper, we study the design of a system that uses work stealing for dynamic load balancing of task-parallel programs executed on hybrid distributed-memory CPU-graphics processing unit (GPU) systems in a global-address space framework. We take into account the unique nature of the accelerator model employed by GPUs, the significant performance difference between GPU and CPU execution as a function of problem size, and the distinct CPU and GPU memory domains. We consider various alternatives in designing a distributed work stealing algorithm for CPU-GPU systems, while taking into account the impact of task distribution and data movement overheads. These strategies are evaluated using microbenchmarks that capture various execution configurations as well as the state-of-the-art CCSD(T) application module from the computational chemistry domain

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2010

Efficient Sparse Matrix-Matrix Multiplication on Heterogeneous High Performance Systems

The efficient implementation of sparse matrix-matrix multiplications on high performance parallel machines poses several challenges: large size of input matrices, compressed representation, density of the output matrices, partitioning and load balancing of matrices that present parts with large differences in density and, thus, in computation times. In this paper we show how, starting from the requirements of such application, we developed a framework that allows its efficient implementation on heterogeneous clusters. We introduce a task based programming model and a runtime supported execution model which provides dynamic load balancing on clusters composed by CPUs and GPUs, allowing better utilization of the system while easing the handling of sparse matrices. Our results show that our solution, which co-designs the application together with the programming model and the runtime system, is able to obtain significant speed-ups due to a virtually perfect load balancing with respect to other programming approaches.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2011

Using a Balun Transformer Combiner for High Power Microwave Experiments

A novel coaxial power combiner design has been duplicated that has distinct advantages over other combiner geometries that can handle high power. This design is being applied to combine four 3 kW power supplies to obtain a 10 kW, 5 MHz system for an ICRF antenna on HSX. In the past, Wilkinson type combiners have had limited application to high power systems because of the lack of non-inductive, high power, 100 Omega balance loads. With this new design, standard 50 Omega dummy loads can be used instead for the balance load. The cost is considerably lower than lumped element combiner designs which are dominated by capacitor costs. At such a relatively low frequency, a 3-dB quarter-wave coupled-line coupler becomes impractically long, and a conventional branch-line hybrid requires 35 Omega-line, which is commercially unavailable. The balun combiner uses less transmission line than a ring hybrid and has good bandwidth characteristics even away from its best line impedance. Theoretical calculations and modeling were performed for line impedances from 65 Omega to 75 Omega. Measurements from a low-power test device show excellent agreement with theory, and construction of the high power system is underway.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2015

On the Impact of Execution Models: A Case Study in Computational Chemistry

Efficient utilization of high-performance computing (HPC) platforms is an important and complex problem. Execution models, abstract descriptions of the dynamic runtime behavior of the execution stack, have significant impact on the utilization of HPC systems. Using a computational chemistry kernel as a case study and a wide variety of execution models combined with load balancing techniques, we explore the impact of execution models on the utilization of an HPC system. We demonstrate a 50 percent improvement in performance by using work stealing relative to a more traditional static scheduling approach. We also use a novel semi-matching technique for load balancing that has comparable performance to a traditional hypergraph-based partitioning implementation, which is computationally expensive. Using this study, we found that execution model design choices and assumptions can limit critical optimizations such as global, dynamic load balancing and finding the correct balance between available work units and different system and runtime overheads. With the emergence of multi- and many-core architectures and the consequent growth in the complexity of HPC platforms, we believe that these lessons will be beneficial to researchers tuning diverse applications on modern HPC platforms, especially on emerging dynamic platforms with energy-induced performance variability.

Execution models, computational chemistry, MPI and↗
DOE OSTI2011

Efficient Support for Matrix Computations on Heterogeneous Multi-core and Multi-GPU Architectures

We present a new methodology for utilizing all CPU cores and all GPUs on a heterogeneous multicore and multi-GPU system to support matrix computations e ciently. Our approach is able to achieve the objectives of a high degree of parallelism, minimized synchronization, minimized communication, and load balancing. Our main idea is to treat the heterogeneous system as a distributed-memory machine, and to use a heterogeneous 1-D block cyclic distribution to allocate data to the host system and GPUs to minimize communication. We have designed heterogeneous algorithms with two di erent tile sizes (one for CPU cores and the other for GPUs) to cope with processor heterogeneity. We propose an auto-tuning method to determine the best tile sizes to attain both high performance and load balancing. We have also implemented a new runtime system and applied it to the Cholesky and QR factorizations. Our experiments on a compute node with two Intel Westmere hexa-core CPUs and three Nvidia Fermi GPUs demonstrate good weak scalability, strong scalability, load balance, and e ciency of our approach.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2010

A Novel Coarsening Method for Scalable and Efficient Mesh Generation

In this paper, we propose a novel mesh coarsening method called brick coarsening method. The proposed method can be used in conjunction with any graph partitioners and scales to very large meshes. This method reduces problem space by decomposing the original mesh into fixed-size blocks of nodes called bricks, layered in a similar way to conventional brick laying, and then assigning each node of the original mesh to appropriate brick. Our experiments indicate that the proposed method scales to very large meshes while allowing simple RCB partitioner to produce higher-quality partitions with significantly less edge cuts. Our results further indicate that the proposed brick-coarsening method allows more complicated partitioners like PT-Scotch to scale to very large problem size while still maintaining good partitioning performance with relatively good edge-cut metric. Graph partitioning is an important problem that has many scientific and engineering applications in such areas as VLSI design, scientific computing, and resource management. Given a graph G = (V,E), where V is the set of vertices and E is the set of edges, (k-way) graph partitioning problem is to partition the vertices of the graph (V) into k disjoint groups such that each group contains roughly equal number of vertices and the number of edges connecting vertices in different groups is minimized. Graph partitioning plays a key role in large scientific computing, especially in mesh-based computations, as it is used as a tool to minimize the volume of communication and to ensure well-balanced load across computing nodes. The impact of graph partitioning on the reduction of communication can be easily seen, for example, in different iterative methods to solve a sparse system of linear equation. Here, a graph partitioning technique is applied to the matrix, which is basically a graph in which each edge is a non-zero entry in the matrix, to allocate groups of vertices to processors in such a way that many of matrix-vector multiplication can be performed locally on each processor and hence to minimize communication. Furthermore, a good graph partitioning scheme ensures the equal amount of computation performed on each processor. Graph partitioning is a well known NP-complete problem, and thus the most commonly used graph partitioning algorithms employ some forms of heuristics. These algorithms vary in terms of their complexity, partition generation time, and the quality of partitions, and they tend to trade off these factors. A significant challenge we are currently facing at the Lawrence Livermore National Laboratory is how to partition very large meshes on massive-size distributed memory machines like IBM BlueGene/P, where scalability becomes a big issue. For example, we have found that the ParMetis, a very popular graph partitioning tool, can only scale to 16K processors. An ideal graph partitioning method on such an environment should be fast and scale to very large meshes, while producing high quality partitions. This is an extremely challenging task, as to scale to that level, the partitioning algorithm should be simple and be able to produce partitions that minimize inter-processor communications and balance the load imposed on the processors. Our goals in this work are two-fold: (1) To develop a new scalable graph partitioning method with good load balancing and communication reduction capability. (2) To study the performance of the proposed partitioning method on very large parallel machines using actual data sets and compare the performance to that of existing methods. The proposed method achieves the desired scalability by reducing the mesh size. For this, it coarsens an input mesh into a smaller size mesh by coalescing the vertices and edges of the original mesh into a set of mega-vertices and mega-edges. A new coarsening method called brick algorithm is developed in this research. In the brick algorithm, the zones in a given mesh are first grouped into fixed size blocks called bricks. These brick are then laid in a way similar to conventional brick laying technique, which reduces the number of neighboring blocks each block needs to communicate. Contributions of this research are as follows: (1) We have developed a novel method that scales to a really large problem size while producing high quality mesh partitions; (2) We measured the performance and scalability of the proposed method on a machine of massive size using a set of actual large complex data sets, where we have scaled to a mesh with 110 million zones using our method. To the best of our knowledge, this is the largest complex mesh that a partitioning method is successfully applied to; and (3) We have shown that proposed method can reduce the number of edge cuts by as much as 65%.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2016

Evaluating the Value of High Spatial Resolution in National Capacity Expansion Models using ReEDS

Power sector capacity expansion models (CEMs) have a broad range of spatial resolutions. This paper uses the Regional Energy Deployment System (ReEDS) model, a long-term national scale electric sector CEM, to evaluate the value of high spatial resolution for CEMs. ReEDS models the United States with 134 load balancing areas (BAs) and captures the variability in existing generation parameters, future technology costs, performance, and resource availability using very high spatial resolution data, especially for wind and solar modeled at 356 resource regions. In this paper we perform planning studies at three different spatial resolutions--native resolution (134 BAs), state-level, and NERC region level--and evaluate how results change under different levels of spatial aggregation in terms of renewable capacity deployment and location, associated transmission builds, and system costs. The results are used to ascertain the value of high geographically resolved models in terms of their impact on relative competitiveness among renewable energy resources.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2013

Energy-Efficient and Comfortable Buildings through Multivariate Integrated Control (ECoMIC)

This project aims to develop an integrated control solution for enhanced energy efficiency and user comfort in commercial buildings. The developed technology is a zone-based control framework that minimizes energy usage while maintaining occupants’ visual and thermal comfort through control of electric lights, motorized venetian blinds and thermostats. The control framework is designed following a modular, scalable and flexible architecture to facilitate easy integration with exiting building management systems. The control framework contains two key algorithms: 1) the lighting load balancing algorithm and 2) the thermostat control algorithm. The lighting load balancing algorithm adopts a model-based closed-loop control approach to determine the optimal electric light and venetian blind settings. It is formulated into an optimization problem with minimizing lighting-related energy consumptions as the objective and delivering adequate task light and preventing daylight glare as the constraints. The thermostat control algorithm is based on a well-established thermal comfort model and formulated as a root-finding problem to dynamically determine the optimal thermostat setpoint for both energy savings and improved thermal comfort. To address building-wide scalability, a system architecture was developed for the zone-based control technology. Three levels of services are defined in the architecture: external services, facility level services and zone level services. The zone-level service includes the control algorithms described above as well as the corresponding interfaces, profiles, sensors and actuators to realize the zone controller. The facility level services connect to the zones through a backbone network, handle supervisory level information and controls, and thus facilitate building-wide scalability. The external services provide communication capability to entities outside of the building for grid interaction and remote access. Various aspects of the developed control technology were evaluated and verified through both simulations and testbed implementations. Simulations coupling a DOE medium office reference building in EnergyPlus building simulation software and a prototype controller in Matlab were performed. During summer time in a mixed-humid climate zone, the simulations revealed reductions of 27% and 42% in electric lighting load and cooling load, respectively, when compared to an advanced base case with daylight dimming and blinds automatically tilted to block direct sun. Two single-room testbeds were established. The testbed at Philips Lighting business building (Rosemont, IL) was designed for quantifying energy performance of integrated controls. This particular implementation achieved 40% and 79% savings on lighting and HVAC energy, respectively, compared to a relatively simple base case operated on predefined schedules. While the resulting energy savings was very encouraging, it should be noted that there may be several caveats associated with it. 1) The test was run during late spring and early summer, and the savings numbers might not be directly used to extrapolate the annual energy savings. 2) Due to the needs for separate control and metering of the small-scale demonstrator within a large building, the HVAC system, hence the corresponding savings, did not represent a typical energy code-compliant design. 3) The light level in the control case was regulated at a particular setpoint, which was lower than then the full-on light level in the base case, and the savings resulted from tuning down the light level to the setpoint was not attributable to the contribution of the developed technology. The testbed at the Lawrence Berkeley National Laboratory (Berkeley, CA) specifically focused on glare control integration, and has demonstrated the feasibility and capability of the glare detection and prevention technique. While the short one-month test in this testbed provided a functional indication of the developed technology, and it would require at least a full solstice-to-solstice cycle to ruinously quantify the performance, which was not possible within the project timeframe. There are certain limitations inherited from the operational assumptions, which could potentially affect the effectiveness and applicability of the developed control technologies. The system takes a typical ceiling-mounting approach for the photosensor locations, and therefore, the control performance relies on proper commissioning or the built-in intelligence of the photosensor for pertinent task light level estimations. For spaces where daylight penetration diminishes significantly deeper into the zone, certain modification to the control algorithms is required to accommodate multiple lighting control subzones and the corresponding sensors for providing a more uniform light level across the entire zone. Integrated control of visual and thermal comfort requires the lighting control zone and thermal control zone to coincide with each other. In other words, the area illuminated by a lighting circuit needs to be the same area served by the thermostat. Thus, the original zoning will potentially constrain the applicability of this technology in retrofitting projects. This project demonstrated the technical feasibility of a zone-based integrated control technology. From the simulation results and testbed implementations, up to 60% lighting energy savings in daylit areas relative to a “no-controls” case can easily be achieved. A 20% reduction of whole building energy consumption is also attainable. In the aspect of occupant comfort, the testbed demonstrated the ability to maintain specified light level on the workplane while promptly mitigate daylight glare 90% of the time. The control system also managed to maintain the thermal environment at a comfortable level 90% of the time. The aspect of system scalability was guaranteed by the system architecture design, based on which the testbeds were instantiated. Analysis on the aspect of economic benefit has yielded an about 6-year payback time for a medium-sized building, including the installation of all hardware and software, such as motorized blinds and LED luminaires. The payback time can be significantly reduced if part of the hardware is already in place for retrofitting projects. It needs to be noted that since the payback analysis was partly based on the testbed performance results, it is constrained by the caveats associated with the testbed implementations. The main uncertainty lies in the contribution from the space conditioning energy savings as it was non-trivial to realistically configure a room-size HVAC system for directly extrapolating whole-building HVAC energy savings. It is recommended to further evaluate the developed technology at a larger scale, where the lighting and HVAC energy consumption can be realistically measured at the building level, to more rigorously quantify the performance potentials.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2019

GraphBLAST: A High-Performance Linear Algebra-based Graph Framework on the GPU

High-performance implementations of graph algorithms are challenging to implement on new parallel hardware such as GPUs, because of three challenges: (1) difficulty of coming up with graph building blocks, (2) load imbalance on parallel hardware, and (3) graph problems having low arithmetic intensity. To address these challenges, GraphBLAS is an innovative, on-going effort by the graph analytics community to propose building blocks based in sparse linear algebra, which will allow graph algorithms to be expressed in a performant, succinct, composable and portable manner. In this paper, we examine the performance challenges of a linear algebra-based approach to building graph frameworks and describe new design principles for overcoming these bottlenecks. Among the new design principles is exploiting input sparsity, which allows users to write graph algorithms without specifying push and pull direction. Exploiting output sparsity allows users to tell the backend which values of the output in a single vectorized computation they do not want computed. Load-balancing is an important feature for balancing work amongst parallel workers. We describe the important load-balancing features for handling graphs with different characteristics. The design principles described in this paper have been implemented in "GraphBLAST", the first open-source linear algebra-based graph framework on GPU targeting high-performance computing. The results show that on a single GPU, GraphBLAST has on average at least an order of magnitude speedup over previous GraphBLAS implementations SuiteSparse and GBTL, comparable performance to the fastest GPU hardwired primitives and shared-memory graph frameworks Ligra and Gunrock, and better performance than any other GPU graph framework, while offering a simpler and more concise programming model.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2011

Exploring Fine-Grained Task-based Execution on Multi-GPU Systems

Many-core Graphics Processing Units (GPUs) have been utilized as the computation engine in many scientific fields due to the high peak performance, cost effectiveness, and the availability of user friendly programming environments, e.g., NVIDIA CUDA. However, the conventional data parallel GPU programming paradigm cannot satisfactorily address issues such as load balancing and GPU resource utilization due to the irregular and unbalanced workload patterns exhibited in some applications. In this paper, we explore the design space of task-based solutions for multi-GPU systems. By employing finer-grained tasks than what is supported in the current CUDA, and allowing task sharing, our solutions enable dynamic load balancing. We evaluate our solutions with a Molecular Dynamics application with different atom distributions (from uniform distribution to highly non-uniform distribution). Experimental results obtained on a 4-GPU system show that, for non-uniform distributed systems, our solutions achieve excellent speedup, and significant performance improvement over other solutions based on the standard CUDA APIs.

99 GENERAL AND MISCELLANEOUS↗
DOE OSTI2020

Scalable Asynchronous Domain Decomposition Solvers

We discuss how parallel implementations of linear iterative solvers generally alternate between phases of data exchange and phases of local computation. Increasingly large problem sizes and more heterogeneous compute architectures make load balancing and the design of low latency network interconnects that are able to satisfy the communication requirements of linear solvers very challenging tasks. In particular, global communication patterns such as inner products become increasingly limiting at scale. We explore the use of asynchronous communication based on one-sided Message Passing Interface primitives in the context of domain decomposition solvers. In particular, a scalable asynchronous two-level Schwarz method is presented. We discuss practical issues encountered in the development of a scalable solver and show experimental results obtained on a state-of-the-art supercomputer system that illustrate the benefits of asynchronous solvers in load balanced as well as load imbalanced scenarios. Using the novel method, we can observe speedups of up to four times over its classical synchronous equivalent.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2013

Accelerating Subsurface Transport Simulation on Heterogeneous Clusters

Reactive transport numerical models simulate chemical and microbiological reactions that occur along a flowpath. These models have to compute reactions for a large number of locations. They solve the set of ordinary differential equations (ODEs) that describes the reaction for each location through the Newton-Raphson technique. This technique involves computing a Jacobian matrix and a residual vector for each set of equation, and then solving iteratively the linearized system by performing Gaussian Elimination and LU decomposition until convergence. STOMP, a well known subsurface flow simulation tool, employs matrices with sizes in the order of 100x100 elements and, for numerical accuracy, LU factorization with full pivoting instead of the faster partial pivoting. Modern high performance computing systems are heterogeneous machines whose nodes integrate both CPUs and GPUs, exposing unprecedented amounts of parallelism. To exploit all their computational power, applications must use both the types of processing elements. For the case of subsurface flow simulation, this mainly requires implementing efficient batched LU-based solvers and identifying efficient solutions for enabling load balancing among the different processors of the system. In this paper we discuss two approaches that allows scaling STOMP's performance on heterogeneous clusters. We initially identify the challenges in implementing batched LU-based solvers for small matrices on GPUs, and propose an implementation that fulfills STOMP's requirements. We compare this implementation to other existing solutions. Then, we combine the batched GPU solver with an OpenMP-based CPU solver, and present an adaptive load balancer that dynamically distributes the linear systems to solve between the two components inside a node. We show how these approaches, integrated into the full application, provide speed ups from 6 to 7 times on large problems, executed on up to 16 nodes of a cluster with two AMD Opteron 6272 and a Tesla M2090 per node.

GPU↗