Code Speed-up Research Articles

High-Performance Computation of the Number of Nested RNA Structures with 3D Parallel Tiled Code

Many current bioinformatics algorithms have been implemented in parallel programming code. Some of them have already reached the limits imposed by Amdahl’s law, but many can still be improved. In our paper, we present an approach allowing us to generate a high-performance code for calculating the number of RNA pairs. The approach allows us to generate parallel tiled code of the maximal dimension of tiles, which for the discussed algorithm is 3D. Experiments carried out by us on two modern multi-core computers, an Intel(R) Xeon(R) Gold 6326 (2.90 GHz, 2 physical units, 32 cores, 64 threads, 24 MB Cache) and Intel(R) i7(11700KF (3.6 GHz, 8 cores, 16 threads, 16 MB Cache), demonstrate a significant increase in performance and scalability of the generated parallel tiled code. For the Intel(R) Xeon(R) Gold 6326 and Intel(R) i7, target code speedup increases linearly with an increase in the number of threads. An approach presented in the paper to generate target code can be used by programmers to generate target parallel tiled code for other bioinformatics codes whose dependence patterns are similar to those of the code implementing the counting algorithm.

Using hardware performance counters to speed up autotuning convergence on GPUs

Nowadays, GPU accelerators are commonly used to speed up general-purpose computing tasks on a variety of hardware. However, due to the diversity of GPU architectures and processed data, optimization of codes for a particular type of hardware and specific data characteristics can be extremely challenging. The autotuning of performance-relevant source-code parameters allows for automatic optimization of applications and keeps their performance portable. Although the autotuning process typically results in code speed-up, searching the tuning space can bring unacceptable overhead if (i) the tuning space is vast and full of poorly-performing implementations, or (ii) the autotuning process has to be repeated frequently because of changes in processed data or migration to different hardware.In this paper, we introduce a novel method for searching generic tuning spaces. The tuning spaces can contain tuning parameters changing any user-defined property of the source code. The method takes advantage of collecting hardware performance counters (also known as profiling counters) during empirical tuning. Those counters are used to navigate the searching process towards faster implementations. The method requires the tuning space to be sampled on any GPU. It builds a problem-specific model, which can be used during autotuning on various, even previously unseen inputs or GPUs. Using a set of five benchmarks, we experimentally demonstrate that our method can speed up autotuning when an application needs to be ported to different hardware or when it needs to process data with different characteristics. We also compared our method to state of the art and show that our method is superior in terms of the number of searching steps and typically outperforms other searches in terms of convergence time.

Parallel Distributed Hydrology Soil Vegetation Model (DHSVM) using global arrays

The Distributed Hydrology Soil Vegetation Model (DHSVM) code was parallelized for distributed memory computers using the Global Arrays (GA) programming model. To analyze parallel performance, DHSVM was used to simulate the hydrology in two river basins of significant size located in the northwest continental United States and southwest Canada at 90 m resolution: the (1) Clearwater (25,000 km2) and (2) Columbia (668,000 km2) River basins. Meteorological forcing applied to both basins was dynamically down-scaled from a regional reanalysis using the Weather Research and Forecasting (WRF) model and read into DHSVM as 2D maps for each time step. Parallel code speedup was significant. Run times for 1-year simulations were reduced by an order of magnitude for both test basins. A maximum parallel speedup of 105 was attained with 480 processors while simulating the Columbia River basin. Speedup was limited by input-dominated tasks, particularly the input of meteorological forcing data.

SMITER: A field-line tracing environment for ITER

Built around the SMARDDA modules for magnetic field-line tracing [IEEE Tr. Plasma Sc. 42 (2014) 1932], the SMITER code package (SMARDDA for ITER) is a new graphical user interface (GUI) framework for power deposition mapping on tokamak plasma-facing components (PFC) in the full 3-D CAD geometry of the machine, taking as input a user-defined specification for parallel heat flux in the scrape-off layer (SOL) and a description of the equilibrium magnetic flux. The software package provides CAD model import and integration with the ITER Integrated Modelling and Analysis Suite (IMAS), parametric CAD components catalogue and modelling, CAD de-featuring for PFC surface extraction, meshing, visualization (using an integrated ParaView module), Python scripting and batch processing, storage in hierarchical data files, with several simulation cases in one study running in parallel and using message passing interface (MPI) for code speed-up. An integrated ParaView module can combine CAD geometry, magnetic field equilibrium, meshes and results for detailed setup analysis and a module is under development for full finite element computation of surface temperatures resulting from the power deposition patterns on 3-D PFCs. The code package has been developed for ITER, but can be deployed for similar modelling of any tokamak. This paper presents and discusses key features of this field-line tracing environment, demonstrates benchmarking against existing field-line tracing code and provides specific examples of power deposition mapping in ITER for different plasma configurations.

Regional Climate Model for the Lower Volga: Parallelization Efficiency Estimation

We have deployed the regional climate model (RCM) RegCM 4.5 for the Lower Volga and adjacent territories with a horizontal spatial resolution of 20 km. The problems of choosing the computational domain in the RCM RegCM version 4.5 are considered. We demonstrate the influence of this factor on the forecast of rainfall distribution in the numerical simulations. The study of rainfall and snowfall is a more demanding test in comparison with temperature or pressure distributions. We investigate dependencies of calculation time, parallel speedup and parallelization efficiency on the number of processes for different multi-core CPUs. Our analysis of the efficiency of parallel implementation of RegCM for various multi-core and multi-processor systems show a strong dependence of the simulation speed on the CPU type. The best effect is achieved when the number of CPU threads and the number of parallel processes are equal. The parallel code speedup is in the range of 1.8 – 11 for different CPUs.

DYNAMIC TILE FREE SCHEDULING FOR CODE WITH ACYCLIC INTER-TILE DEPENDENCE GRAPHS

Free scheduling is a task ordering technique under which instructions are executed as soon as their operands become available. Coarsening the grain of computations under the free schedule, by means of using groups of loop nest statement instances (tiles) in place of single statement instances, increases the locality of data accesses and reduces the number of synchronization events, and as a consequence improves program performance. The paper presents an approach for code generation allowing for the free schedule for tiles of arbitrarily nested affine loops at run-time. The scope of the applicability of the introduced algorithms is limited to tiled loop nests whose inter-tile dependence graphs are cycle-free. The approach is based on the Polyhedral Model. Results of experiments with the PolyBench benchmark suite, demonstrating significant tiled code speed-up, are discussed.

Efficient code simulation strategies for B2-EIRENE

Plasma edge simulation codes are crucial for the interpretation of present experiments as well as for the assessment of new concepts and next generation nuclear fusion devices. These codes are most often based on a combined Finite Volume (FV) / Monte Carlo (MC) numerical approach to simulate coupled plasma and neutral particle transport. In this paper we apply recently derived error reduction analysis to assess numerical errors in a partially detached ITER case. We show that also for this strongly coupled FV/MC simulation case, statistical averaging over iterations provides an accurate means to achieve well reproducible and statistically accurate results. Moreover, the error reduction analysis with respect to numerical parameters provide a framework to achieve increased accuracy for a given computational cost. We also show how significant code speed-up can be achieved for a desirable accuracy compared to presently used simulation strategies.

Optimizations of the energy grid search algorithm in continuous-energy Monte Carlo particle transport codes

In this work we propose, implement, and test various optimizations of the typical energy grid-cross section pair lookup algorithm in Monte Carlo particle transport codes. The key feature common to all of the optimizations is a reduction in the length of the vector of energies that must be searched when locating the index of a particle’s current energy. Other factors held constant, a reduction in energy vector length yields a reduction in CPU time. The computational methods we present here are physics-informed. That is, they are designed to utilize the physical information embedded in a simulation in order to reduce the length of the vector to be searched. More specifically, the optimizations take advantage of information about scattering kinematics, neutron cross section structure and data representation, and also the expected characteristics of a system’s spatial flux distribution and energy spectrum. The methods that we present are implemented in the OpenMC Monte Carlo neutron transport code as part of this work. The gains in computational efficiency, as measured by overall code speedup, associated with each of the optimizations are demonstrated in both serial and multithreaded simulations of realistic systems. Depending on the system, simulation parameters, and optimization method employed, overall code speedup factors of 1.2–1.5, relative to the typical single-nuclide binary search algorithm, are routinely observed.

Cosmological calculations on the GPU

Abstract Cosmological measurements require the calculation of nontrivial quantities over large datasets. The next generation of survey telescopes will yield measurements of billions of galaxies. The scale of these datasets, and the nature of the calculations involved, make cosmological calculations ideal models for implementation on graphics processing units (GPUs). We consider two cosmological calculations, the two-point angular correlation function and the aperture mass statistic, and aim to improve the calculation time by constructing code for calculating them on the GPU. Using CUDA, we implement the two algorithms on the GPU and compare the calculation speeds to comparable code run on the CPU. We obtain a code speed-up of between 10 and 180 × faster, compared to performing the same calculation on the CPU. The code has been made publicly available. GPUs are a useful tool for cosmological calculations, even for datasets the size of current surveys, allowing calculations to be made one or two orders of magnitude faster.

Full-tungsten plasma edge simulations with SOLPS

We review different approaches available for plasma edge modeling when including tungsten impurities. We focus on the bundled charged states method, and perform a sensitivity study, against a full fluid treatment, for a benchmark ASDEX-Upgrade case. We consider several indicators, such as net wall erosion/deposition rates and plasma power balance. In general, differences are modest, but material migration rates are slightly underestimated with bundles versus the full treatment. At the same time, radiated power and heat flux to the divertors can vary significantly from one bundling scheme to the other. It therefore appears that bundling of several ionic stages into single atomic “species”, while providing a marked code speed-up (by reducing the number of equations by up to a factor 4), must be used with care when making divertor behavior predictions.

Advanced solution algorithms for transient multidimensional thermohydraulic flow problems in complex geometries with the programme COMMIX-2/KfK

The computer programme COMMIX-2 describes steady state and transient multidimensional single- and two-phase fluid flows with heat transfer in nuclear reactor components and multicomponent systems. Originally from the Argonne National Laboratory, the code has been further developed at the Kernforschungszentrum Karlsruhe. The original Point-SOR iterative method for the solution of a Poisson-like equation describing the pressure distribution in the fluid as well as the transport of enthalpy and turbulent quantities has been complemented with iterative and direct line- and block-methods. None of the newly implemented methods is original in itself but their implementation into the computer code, which can describe the most general shapes of definition domains, gave a code speed-up by a factor of 2–5, depending on the problem treated. The code capabilities are assessed by the calculation of a benchmark problem involving the numerical simulation of thermal buoyancy phenomena at a pipe/plenum interface.

Measuring the Parallelism Available for Very Long Instruction Word Architectures

Long instruction word architectures, such as attached scientific processors and horizontally microcoded CPU's, are a popular means of obtaining code speedup via fine-grained parallelism. The falling cost of hardware holds out the hope of using these architectures for much more parallelism. But this hope has been diminished by experiments measuring how much parallelism is available in the code to start with. These experiments implied that even if we had infinite hardware, long instruction word architectures could not provide a speedup of more than a factor of 2 or 3 on real programs.