Speedup Prediction Research Articles

Decision support for GPU acceleration by predicting energy savings and programming effort

As the number of heterogeneous embedded systems used in IoT applications increases, there is a lack of software tools to assist developers to meet the challenge of reducing energy consumption. Indeed, there are only few performance prediction tools for heterogeneous systems in the literature and they typically focus on the prediction of speedup by acceleration. In this work, we propose a methodology for analyzing CPU applications in order to estimate the potential Energy gains by offloading a piece of code on an embedded GPU. The proposed methodology provides several features beyond the state of the art of existing predictors, including the combination of static analysis and dynamic instrumentation approaches and the prediction of the programming effort of developing the CUDA kernel of a CPU code, using advanced metrics. The methodology is supported by a tool-flow and it is demonstrated and evaluated on modern heterogeneous embedded systems (Nvidia), where shows classification accuracy above 75%. The results show that the proposed methodology can assist application developers in the early design choice of investing effort to acceleration considering the expected Energy Savings and the Effort required to develop acceleration-specific code.

Effectiveness of coarse graining degree and speedup on the dynamic properties of homopolymer.

Evaluation of effective coarse graining (CG) degree and reasonable speedup relative to all-atomistic (AA) model was conducted to provide a basis for building appropriate larger-scale model. The reproducibility of atomistic conformation and temperature transferability both act as the analysis criteria to resolve the maximum acceptable CG degree. Taking short- and long time spans into account simultaneously in the estimation of computational speedup, a dynamic scaling factor is accessible in fitting mean squared displacement ratio of CG to AA as an exponential function. Computing loss in parallel running is an indispensable component in acceleration, which was also added in the evaluation. Subsequently, a quantified prediction of CG speedup arises as a multiplication of dynamic scaling factor, computing loss, time step, and the square of reduction in the number of degrees of freedom. Polyethylene oxide was adopted as a reference system to execute the direct Boltzmann inversion and iterative Boltzmann inversion. Bonded and non-bonded potentials were calculated in CG models with 1~4 monomers per bead. The effective CG degree was determined as two at the most with a speedup of four orders magnitude over AA in this study. Determination of effectiveness CG degree and the corresponding speedup prediction provide available tools in larger spatiotemporal-scale calculations.

Vectorization cost modeling for NEON, AVX and SVE

Compiler optimization passes employ cost models to determine if a code transformation will yield performance improvements. When this assessment is inaccurate, compilers apply transformations that are not beneficial, or refrain from applying ones that would have improved the code. We analyze the accuracy of the cost models used in LLVM’s and GCC’s vectorization passes for three different instruction set architectures, including both traditional SIMD architectures with a defined fixed vector register size (AVX2 and NEON), and novel instruction set with scalable vector size (SVE). In general, speedup is over-estimated, resulting in mispredictions and a weak to medium correlation between predicted and actual performance gain. We therefore propose a novel cost model that is based on a code’s intermediate representation with refined memory access pattern features. Using linear regression techniques, this platform independent model is fitted to an AVX2 and a NEON hardware, as well as an SVE simulator. Results show that the fitted model significantly improves the correlation between predicted and measured speedup (AVX2: +52% for training data, +13% for validation data), reduces the average error of the speedup prediction (SVE: -43% for training data, -36% for validation data), as well as the number of mispredictions (NEON: -88% for training data, -71% for validation data) for more than 80 code patterns.

Wind speed measurements and predictions over Belmont Hill, Wellington, New Zealand

The aim of the research described in this paper is to reduce the vulnerability of New Zealand's built infrastructure to wind damage through provision of improved wind speed prediction procedures for wind flow over complex terrain that could ultimately be incorporated into the AS/NZS 1170.2 loading standard. This paper presents results from a specific experiment which compares measured wind speedups (neutral atmosphere) over the rugged Belmont Hill region with wind speedups estimated from AS/NZS 1170.2 by two organisations, and from six other international wind-loading standards, through Computational Fluid Dynamics (Gerris and WAsP) modelling, and through wind-tunnel modelling. It was found that computer modelling with the large-eddy simulation CFD code Gerris and wind-tunnel measurements, agree remarkably well with the observations and differentiate considerably better between areas exposed or sheltered by the local terrain features compared to applying the CFD code WAsP (4.1), or applying the AS/NZS 1170.2 wind-loading standard. Furthermore, it was found that the AS/NZS1170.2 predictions differed significantly when carried out by different organisations. The wind speedup prediction methods from the six other international standards all use similar approaches and thus the predictions from all of them were quite similar, and did not agree particularly well with the field measurements.

Comparison of Speed-Up Over Hills Derived from Wind-Tunnel Experiments, Wind-Loading Standards, and Numerical Modelling

We evaluate the accuracy of the speed-up provided in several wind-loading standards by comparison with wind-tunnel measurements and numerical predictions, which are carried out at a nominal scale of 1:500 and full-scale, respectively. Airflow over two- and three-dimensional bell-shaped hills is numerically modelled using the Reynolds-averaged Navier–Stokes method with a pressure-driven atmospheric boundary layer and three different turbulence models. Investigated in detail are the effects of grid size on the speed-up and flow separation, as well as the resulting uncertainties in the numerical simulations. Good agreement is obtained between the numerical prediction of speed-up, as well as the wake region size and location, with that according to large-eddy simulations and the wind-tunnel results. The numerical results demonstrate the ability to predict the airflow over a hill with good accuracy with considerably less computational time than for large-eddy simulation. Numerical simulations for a three-dimensional hill show that the speed-up and the wake region decrease significantly when compared with the flow over two-dimensional hills due to the secondary flow around three-dimensional hills. Different hill slopes and shapes are simulated numerically to investigate the effect of hill profile on the speed-up. In comparison with more peaked hill crests, flat-topped hills have a lower speed-up at the crest up to heights of about half the hill height, for which none of the standards gives entirely satisfactory values of speed-up. Overall, the latest versions of the National Building Code of Canada and the Australian and New Zealand Standard give the best predictions of wind speed over isolated hills.

Using sequential runtime distributions for the parallel speedup prediction of SAT local search

AbstractThis paper presents a detailed analysis of the scalability and parallelization of local search algorithms for the Satisfiability problem. We propose a framework to estimate the parallel performance of a given algorithm by analyzing the runtime behavior of its sequential version. Indeed, by approximating the runtime distribution of the sequential process with statistical methods, the runtime behavior of the parallel process can be predicted by a model based on order statistics. We apply this approach to study the parallel performance of two SAT local search solvers, namely Sparrow and CCASAT, and compare the predicted performances to the results of an actual experimentation on parallel hardware up to 384 cores. We show that the model is accurate and predicts performance close to the empirical data. Moreover, as we study different types of instances (random and crafted), we observe that the local search solvers exhibit different behaviors and that their runtime distributions can be approximated by two types of distributions: exponential (shifted and non-shifted) and lognormal.

Experimental and theoretical speedup prediction of MPI-based applications

Prediction of speedup obtained from parallelization plays an important role in converting serial applications into parallel ones. Several parameters affect the execution time of an application. In this paper we experimentally and theoretically study the effect of some of these parameters on the execution time of Message Passing Interface based applications.

Analysis of speedup as function of block size and cluster size for parallel feed-forward neural networks on a Beowulf cluster.

The performance of feed-forward neural networks trained with the backpropagation algorithm on a dedicated Beowulf cluster is analyzed. The concept of training set parallelism is applied. A new model for run time and speedup prediction is developed. With the model the speedup and efficiency of one iteration of the neural networks can be estimated as a function of block size and cluster size. The model is applied to three example problems representing different applications and network architectures. The estimation of the model has a higher accuracy than traditional methods for run time estimation and can be efficiently calculated. Experiments show that speedup of one iteration does not necessarily translate to a shorter training time toward a given error level. To overcome this problem a heuristic extension to training set parallelism called weight averaging is developed. The results show that training in parallel should only be done on clusters with high performance network connections or a multiprocessor machine. A rule of thumb is given for how much network performance of the cluster is needed to achieve speedup of the training time for a neural network.

An analysis of the recycle dynamics of a high pressure production process

Control schemes have been developed for a continuous pilot plant for vitamin production using the approach recommended by Luyben. The plant was simulated using Speedup and accurate physical property prediction. The approach was modified to fit the plant but it was confirmed that the basic premise of requiring one stream in the recycle loop under flow control was confirmed.

Parallel pivotal algorithm for solving the linear complementarity problem

We propose a parallel implementation of the classical Lemke's algorithm for solving the linear complementarity problem. The algorithm is designed for a loosely coupled network of computers which is characterized by relatively high communication costs. We provide an accurate prediction of speedup based on a simple operation count. The algorithm produces speedup nearp, wherep is the number of processors, when tested on large problems as demonstrated by computational results on the CRYSTAL token-ring multicomputer and the Sequent Balance 21000 multiprocessor.