Table 2 - available via license: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International
Content may be subject to copyright.
Source publication
Trade-off between the cost-efficiency of powerful computational accelerators and the increasing energy needed to perform numerical tasks can be tackled by implementation of algorithms on the Intel Multiple Integrated Cores (MIC) architecture. The best performance of the algorithms requires the use of appropriate optimization and parallelization app...
Citations
... We use computer simulations of the atmospheric parameters that define the thermal comfort by Advanced Research Weather Forecast Model WRF ARW version 3.9 (Skamarock et al. 2008). The simulations were performed on the HPC System "Avitohol" at the Institute of Information and Communication Technologies at the Bulgarian Academy of Sciences (Atanassov et al. 2017;Karaivanova et al. 2022) and at the NESTUM cluster. They were performed on four nested domains ( Fig. 1) for 2017 with an output frequency of 1 hour. ...
According to future climate projections, the expected thermal stress environmental conditions will get worse if the authorities do not apply appropriate measures for mitigation and adaptation to these changes. The health issues concerning air pollution and extreme temperatures have assumed great importance in recent years. The objective of this study is to estimate the thermal comfort in two of the biggest cities in Bulgaria-Sofia and Varna and their surroundings for the year 2017 by biometeorological indexes. We use computer simulations of the atmospheric parameters that define the thermal comfort indexes, by Advanced Research Weather Forecast Model WRF ARW version 3.9. We performed the simulations on four domains for 2017 with an output frequency of 1 hour. The outermost domain has a horizontal resolution of 9 km and encompasses the Balkan Peninsula. It uses initial and boundary conditions from the 0.25-degree NCEP Final Operational Model Global Tropospheric Analyses datasets with a time-frequency of 6 hours. The estimation of the thermal comfort conditions is performed with characteristics called indexes. The differences in the number of cases between the indexes are due to the specific definitions and the meteorological factors that each of them takes into account. Some of these characteristics have applications depending on the specific thermal conditions.
... In that way the "typical" annual estimates were obtained. The calculations were carried out on the supercomputer system "Avithol" situated at the Institute of Information and Communication Technologies at the Bulgarian Academy of Sciences (IIKT-BAS), as the simulations were organized in different jobs (Atanassov et al. 2017, Karaivanova et al. 2022. ...
Bulgaria has developed national emission reduction strategies for the period from 2020 to 2029 and the years after 2030, in accordance with EU Directive 2016/2284. Our fundamental aim in this study is to assess the effects of the strategy on the PM near surface concentrations in Bulgaria. All the simulation was done by the modeling system U.S. Environmental Protection Agency (US EPA) Models-3 for 2008 to 2014 period and with 9 km horizontal grid resolution for the selected region – Bulgaria. The meteorological background that was used is with 1°x1° resolution from the National Centers for Environmental Prediction (NCEP) Global Analysis Data. There are 5 emission scenarios structured: 2005 emissions (reference period), 2020–2029 emissions projected with existing measures (WEM) and with additional measures (WAM), projected after 2030 WEM and WAM emissions, as parallel calculations were performed with all of the scenarios. Making parallels between the concentrations, with different scenarios simulated, gives the possibility to evaluate the national emission reduction strategies’ effect.
... Acknowledgement_List.pdf. The extensive numerical calculations were done on the Avitohol supercomputer that is described in [53]. ...
Introduction
The Amyloid/Tau/Neurodegeneration (ATN) framework was proposed to identify the preclinical biological state of Alzheimer’s disease (AD). We investigated whether ATN phenotype can be predicted using routinely collected research cohort data.
Methods
927 EPAD LCS cohort participants free of dementia or Mild Cognitive Impairment were separated into 5 ATN categories. We used machine learning (ML) methods to identify a set of significant features separating each neurodegeneration-related group from controls (A-T-(N)-). Random Forest and linear-kernel SVM with stratified 5-fold cross validations were used to optimize model whose performance was then tested in the ADNI database.
Results
Our optimal results outperformed ATN cross-validated logistic regression models by between 2.2% and 8.3%. The optimal feature sets were not consistent across the 4 models with the AD pathologic change vs controls set differing the most from the rest. Because of that we have identified a subset of 10 features that yield results very close or identical to the optimal.
Discussion
Our study demonstrates the gains offered by ML in generating ATN risk prediction over logistic regression models among pre-dementia individuals.
... Acknowledgement_List.pdf. The extensive numerical calculations were done on the Avitohol supercomputer that is described in [53]. ...
There is a strong case for de‐risking neurodegenerative agent development through highly informative experimental medicine studies early in the disease process. These types of studies are dependent on a research infrastructure that includes volunteer registries holding highly granular phenotypic and genotypic data to allow stratified study selection. Examples of such registries include the Brain Health Registry, Great Minds and PROTECT cohorts which rely on remote cognitive, self‐reported medical history and genetic data. This requires the development of effective algorithms to predict the presence of preclinical dementia pathology. In this study we sought to address this need by building a machine learning (ML) ATN risk prediction algorithm which incorporates data typically collected in such registries. To build a ML algorithm that is validated against an existing regression‐based model (Calvin et al. 2020), we used the EPAD LCS cohort (V1500.0). We excluded participants with 1) known diagnosis of dementia or Mild Cognitive Impairment or Clinical Dementia Rating scale ≥ 0.5 and 2) no cerebrospinal fluid biomarkers. Participants were categorised into 5 ATN categories: (i) Normal AD biomarkers: A−T−(N)−; (ii) Alzheimer’s pathologic change: A+T−(N)−; (iii) Alzheimer’s disease: A+T+(N)±; (iv) Alzheimer’s and concomitant non‐Alzheimer’s pathologic change: A+T−(N)+; (v) Non‐AD pathologic change: A−T ± (N)+; A−T+(N)−. Using a Weight of Evidence and Information Value method we identified 13 significant features for testing differences between each of the four neurodegeneration‐related groups vs. controls (A‐T‐N‐). Random Forest and XGBoost with 5‐fold cross validations were used to optimise the Area Under the Curve (AUC) metric. The study dataset included 927 individuals. Our optimal results outperformed the regression models in the Calvin et al. 2020 paper by between 2 and 12%. The optimal feature sets were not consistent across the 4 models with the A+T−(N)+ vs A−T−(N)− differing the most from the rest. Our study demonstrates the gains offered by ML in generating ATN risk prediction over logistic regression models among pre‐dementia individuals. The reliance of the model on variables that can be collected remotely demonstrates its utility for research registers. An openly available version of the ML algorithm for use by research registries is under development.
... Particle transport plays an important role in modeling many physical phenomena and engineering problems. Particle transport theory has been applied in (Atanassov et al., 2017) astrophysics (Chandrasekhar, 2013), nuclear physics (Marchuk and Lebedev, 1986), medical radiotherapy (Bentel, 2009), and many other fields. The particle transport equation (Boltzmann equation) is a mathematical physics equation describing the particle transport process, and its solution algorithm has always been the key to research in this field. ...
... The existing commonly used solutions are divided into two categories, one is the deterministic methods for solving algebraic equations through discrete space, including spherical harmonic method (Marshak, 1947), discrete ordinates method (Carlson, 1955), etc. The other is the stochastic methods, for instance, the Monte Carlo method (Eckhardt, 1987), which simulates particle space using probability theory (Atanassov et al., 2017). With the development of science and technology, simulations of particle transport is more and more demanding of precision and realtimeness. ...
Scalable parallel algorithm for particle transport is one of the main application fields in high-performance computing. Discrete ordinate method (S n ) is one of the most popular deterministic numerical methods for solving particle transport equations. In this paper, we introduce a new method of large-scale heterogeneous computing of one energy group time-independent deterministic discrete ordinates neutron transport in 3D Cartesian geometry (Sweep3D) on Tianhe-2A supercomputer. In heterogeneous programming, we use customized Basic Communication Library (BCL) and Accelerated Computing Library (ACL) to control and communicate between CPU and the Matrix2000 accelerator. We use OpenMP instructions to exploit the parallelism of threads based on Matrix 2000. The test results show that the optimization of applying OpenMP on particle transport algorithm modified by our method can get 11.3 times acceleration at most. On Tianhe-2A supercomputer, the parallel efficiency of 1.01 million cores compared with 170 thousand cores is 52%.
... The extensive numerical calculations were performed on the HPC facility Avitohol of IICT-BAS, described in [1]. ...
In this work we investigate the properties of the statistical power of the Benjamini-Hochberg procedure. That procedure provides a widely used method for controlling the False Discovery Rate in a multiple comparison setup. We show that in many cases the statistical power of the p-values adjusted by the Benjamini-Hochberg procedure could be approximated by a normal distribution with both its mean and standard deviation having an exponential fit and convergence when the number of tests increase. As a result one could estimate the probability that the power belongs in a predetermined interval in a very computationally efficient way using only simulations for several values of the number of tests parameter. The fit for the mean also supports the conjecture that the power of the test decreases to the limit power, which is known to exist, with increasing the number of tests. The latter is a very favourable observation from practical perspective and we try to offer partially rigorous explanation why the monotonicity is present.
... At the end of this section we show the execution time breakdown for the Monte Carlo sparse inverse preconditioner and six test matrices from The University of Florida Sparse Matrix Collection [30] and Matrix market [6], described in Table 1. Figs 1 and 2 show the large communication time which was the motivation for our developments presented in this paper. ...
The tightened energy requirements when designing state-of-the-art high performance computing systems lead to the increased use of computational accelerators. Intel introduced the Many Integrated Core (MIC) architecture for their line of accelerators and successfully competes with NVIDIA on basis of price/performance and ease of development. Although some codes may be ported successfully to Intel MIC architecture without significant modifications, in order to achieve optimal performance one has to make the best use of the vector processing capabilities of the architecture. In this work we present our implementation of Quasi-Monte Carlo methods for matrix computations specifically optimised for the Intel Xeon Phi accelerators. To achieve optimal parallel efficiency we make use of both MPI and OpenMP.
When analyzing biomedical data, researchers often need to apply the same statistical test or procedure to many variables resulting in a multiple comparisons setup. A portion of the tests are statistically significant, their unadjusted p-values form a spike near the origin and as such they can be modeled by a suitable Beta distribution. The unadjusted p-values of the non-significant tests are drawn form a uniform distribution in the unit interval. Therefore the set of all unadjusted p-values can be represented by a beta-uniform mixture model. Finding the parameters of that model plays an important role in estimating the statistical power of the subsequent Benjamini-Hochberg correction for multiple comparisons. To empirically investigate the properties of some parameter estimation procedures we carried out a series of computationally intensive numerical simulations on a high-performance computing facility. As a result of these simulations, in this article we have identified the overall optimal method for estimating the mixture parameters. We also show an asymptotic property of one of the parameter estimates.
Missing data is a common problem when analysing real-world data from many different research fields such as biostatistics, sociology, economics etc. Three types of missing data are typically defined: missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). Ignoring observations with missingness could lead to serious bias and inefficiency, especially when the number of such cases is large compared to the sample size. One popular technique for solving the missing data issue is multiple imputation (MI). There are two general approaches to MI. One is joint modelling which draws missing values simultaneously for all incomplete variables from a multivariate distribution. The other is the fully conditional specification (FCS, also known as MICE), which imputes variables one at a time from a series of univariate conditional distributions. For each incomplete variable FCS draws from a univariate density conditional on the other variables included in the imputation model. In this work we define a computationally efficient numerical simulation framework for data generation and evaluation of different imputation methods. We consider different FCS imputation methods along with traditional ones under different scenarios for the parameters of the models - percentage of missingness, data dimensionality, different combination of categorical and numerical predictors and different correlation between the covariates. Our results are based on synthetic data generated on HPC cluster and show the optimal imputation methods in the different cases according to two scoring techniques.
