Multi-level Parallelism Research Articles

Multilevel Parallelism for the Exploration of Large-Scale Graphs

We present the most recent release of our parallel implementation of the BFS and BC algorithms for the study of large scale graphs. Although our reference platform is a high-end cluster of new generation Nvidia GPUs and some of our optimizations are CUDA specific, most of our ideas can be applied to other platforms offering multiple levels of parallelism. We exploit multi level parallel processing through a hybrid programming paradigm that combines highly tuned CUDA kernels, for the computations performed by each node, and explicit data exchange through the Message Passing Interface (MPI), for the communications among nodes. The results of the numerical experiments show that the performance of our code is comparable or better with respect to other state-of-the-art solutions. For the BFS, for instance, we reach a peak performance of 200 Giga Teps on a single GPU and 5.5 Terateps on 1024 Pascal GPUs. We release our source codes both for reproducing the results and for facilitating their usage as a building block for the implementation of other algorithms.

Parallel Multilevel Cell STT-MRAMs for Optimized Area Energy and Read–Write Operations

Spin-transfer torque (STT) magnetic random access memory (MRAM) is considered as a promising candidate for the next generation embedded memory applications. The advent of a perpendicular magnetic anisotropy-based magnetic tunnel junction (PMTJ) device has made the STT-MRAM viable for low-power high-density on-chip memory architectures. However, intense research efforts are still required to optimize the single-level cell and multilevel cell (MLC) STT-MRAMs in terms of area, energy, and read–write operations at ultralow power supply voltages. In this paper, a parallel MLC (pMLC) STT-MRAM (20–50 nm) is proposed to provide smaller footprint area, energy efficient, and optimized read–write operations with very low failure probabilities at the supply voltage of 1 V. These designs utilize high- ${k}$ gate dielectric gate-all-around silicon nanowire-based vertical NMOS transistor as a switching device. The PMTJ device is modeled using Verilog-A. The HSPICE simulation results affirm the suitability of pMLC designs, e.g., 3-bit pMLC design exhibits, switching error probability much less than 10−9, and average write energy of <2 pJ with minimum cell-area at the pulsewidth of 10 ns.

Current correlations for the transport of interacting electrons through parallel quantum dots in a photon cavity

We calculate the current correlations for the steady-state electron transport through multi-level parallel quantum dots embedded in a short quantum wire, that is placed in a non-perfect photon cavity. We account for the electron–electron Coulomb interaction, and the para- and diamagnetic electron–photon interactions with a stepwise scheme of configuration interactions and truncation of the many-body Fock spaces. In the spectral density of the temporal current–current correlations we identify all the transitions, radiative and non-radiative, active in the system in order to maintain the steady state. We observe strong signs of two types of Rabi oscillations.

On the achievement of high fidelity and scalability for large‐scale diagonalizations in grid‐based DFT simulations

AbstractRecent advance in high performance computing (HPC) resources has opened the possibility to expand the scope of density functional theory (DFT) simulations toward large and complex molecular systems. This work proposes a numerically robust method that enables scalable diagonalizations of large DFT Hamiltonian matrices, particularly with thousands of computing CPUs (cores) that are usual these days in terms of sizes of HPC resources. The well‐known Lanczos method is extensively refactorized to overcome its weakness for evaluation of multiple degenerate eigenpairs that is the substance of DFT simulations, where a multilevel parallelization is adopted for scalable simulations in as many cores as possible. With solid benchmark tests for realistic molecular systems, the fidelity of our method are validated against the locally optimal block preconditioned conjugated gradient (LOBPCG) method that is widely used to simulate electronic structures. Our method may waste computing resources for simulations of molecules whose degeneracy cannot be reasonably estimated. But, compared to LOBPCG method, it is fairly excellent in perspectives of both speed and scalability, and particularly has remarkably less (&lt; 10%) sensitivity of performance to the random nature of initial basis vectors. As a promising candidate for solving electronic structures of highly degenerate systems, the proposed method can make a meaningful contribution to migrating DFT simulations toward extremely large computing environments that normally have more than several tens of thousands of computing cores.

Multilevel parallelization for simulating compressible turbulent flows on most kinds of hybrid supercomputers

The paper describes a multilevel MPI+OpenMP+OpenCL parallelization approach that provides complete portability across a wide range of hybrid supercomputer architectures. A parallel CFD algorithm for heterogeneous computing of turbulent flows is presented. It simulates the compressible Navier–Stokes equations using a cell-centered finite-volume method with polynomial reconstruction on unstructured hybrid meshes. A two-level partitioning is used for the workload distribution among computing devices of hybrid nodes. The overlap of communications and computations hides the data transfer expenses. The scalability is tested on various HPC systems including a fat node with 8 GPUs and supercomputers using up to 320 GPUs. Comparison of performance is presented for multicore CPUs, Intel Xeon Phi, various GPUs of AMD and NVIDIA. The heterogeneous execution using CPUs and GPUs is studied in detail.

Multi-level dataset decomposition for parallel frequent itemset mining on a cluster of personal computers

Frequent Itemset mining is time consuming for large datasets. Many parallel frequent itemset mining algorithms have been proposed to speed up the mining process. This paper presents a parallel frequent itemset mining algorithm on a cluster of personal computers. To facilitate parallel frequent itemset mining, we use prefix path based method to decompose a transactional dataset into its frequent 1-itemset sub-datasets. We called the parallel frequent itemset mining algorithm based on the frequent 1-itemset sub-dataset decomposition the single-level parallel frequent itemset mining algorithm (SLPFIM) in our PC cluster platform. To mitigate the bottleneck caused by time-consuming 1-itemset sub-datasets, we propose a multi-level parallel frequent itemset mining (MLPFIM) algorithm to further decompose the time-consuming 1-itemset sub-datasets into their corresponding sub-sub-datasets. The fine granule of the sub-sub-datasets enhances the load balancing in parallel frequent itemset mining. The experimental results showed that the SLPFIM offered a maximum of 11.9x speedup over the non-parallel execution of the FP-Growth algorithm while the MLPFIM achieved a maximum of 23.1x speedup over the non-parallel execution of the FP-Growth algorithm. The experimental results also showed that the MLPFIM offered a maximum of 2.14x speedup over the SLPFIM.

Nuclide depletion capabilities in the Shift Monte Carlo code

A new depletion capability has been developed in the Exnihilo radiation transport code suite. This capability enables massively parallel domain-decomposed coupling between the Shift continuous-energy Monte Carlo solver and the nuclide depletion solvers in ORIGEN to perform high-performance Monte Carlo depletion calculations. This paper describes this new depletion capability and discusses its various features, including a multi-level parallel decomposition, high-order transport-depletion coupling, and energy-integrated power renormalization. Several test problems are presented to validate the new capability against other Monte Carlo depletion codes, and the parallel performance of the new capability is analyzed.

Parallel Multiclass Support Vector Machine for Remote Sensing Data Classification on Multicore and Many-Core Architectures

Support Vector Machine (SVM) is a classification method that has been widely used in the domain of remote sensing for decades. Although SVM-based classification method achieves good performance for classification accuracy in many studies, it can become very time-consuming in some remote sensing applications such as hyperspectral image classification or large-scale land cover mapping. To improve the efficiency for SVM training and classification in remote sensing applications, we designed and implemented a highly efficient multiclass support vector machine (MMSVM) for × 86-based multicore and many-core architectures such as the Ivy Bridge CPUs and the Intel Xeon Phi coprocessor (MIC) based on our previous MIC-SVM library. Various analysis methods and optimization strategies are employed to fully utilize the multilevel parallelism of our studied architectures. We select several real-world remote sensing datasets to evaluate the performance of our proposed MMSVM. Compared with the widely used serial LIBSVM, our MMSVM achieves 6.3–31.1 $\times$ (in training) and 4.9–32.2 $\times$ (in classification) speedups on MIC, and 6.9–14.9 $\times$ (in training) and 5.5–22.1 $\times$ (in classification) speedups on the Ivy Bridge CPUs. We also conduct a performance comparison analysis on different platforms and provide some ideas on how to select the most suitable architecture for specific remote sensing classification problems in order to achieve the best performance.

A technique to automatically determine Ad-hoc communication patterns at runtime

Abstract Current High Performance Computing (HPC) systems are typically built as interconnected clusters of shared-memory multicore computers. Several techniques to automatically generate parallel programs from high-level parallel languages or sequential codes have been proposed. To properly exploit the scalability of HPC clusters, these techniques should take into account the combination of data communication across distributed memory, and the exploitation of shared-memory models. In this paper, we present a new communication calculation technique to be applied across different SPMD (Single Program Multiple Data) code blocks, containing several uniform data access expressions. We have implemented this technique in Trasgo, a programming model and compilation framework that transforms parallel programs from a high-level parallel specification that deals with parallelism in a unified, abstract, and portable way. The proposed technique computes at runtime exact coarse-grained communications for distributed message-passing processes. Applying this technique at runtime has the advantage of being independent of compile-time decisions, such as the tile size chosen for each process. Our approach allows the automatic generation of pre-compiled multi-level parallel routines, libraries, or programs that can adapt their communication, synchronization, and optimization structures to the target system, even when computing nodes have different capabilities. Our experimental results show that, despite our runtime calculation, our approach can automatically produce efficient programs compared with MPI reference codes, and with codes generated with auto-parallelizing compilers.

Accelerated parallel algorithm for gene network reverse engineering

BackgroundThe Algorithm for the Reconstruction of Accurate Cellular Networks (ARACNE) represents one of the most effective tools to reconstruct gene regulatory networks from large-scale molecular profile datasets. However, previous implementations require intensive computing resources and, in some cases, restrict the number of samples that can be used. These issues can be addressed elegantly in a GPU computing framework, where repeated mathematical computation can be done efficiently, but requires extensive redesign to apply parallel computing techniques to the original serial algorithm, involving detailed optimization efforts based on a deep understanding of both hardware and software architecture.ResultHere, we present an accelerated parallel implementation of ARACNE (GPU-ARACNE). By taking advantage of multi-level parallelism and the Compute Unified Device Architecture (CUDA) parallel kernel-call library, GPU-ARACNE successfully parallelizes a serial algorithm and simplifies the user experience from multi-step operations to one step. Using public datasets on comparable hardware configurations, we showed that GPU-ARACNE is faster than previous implementations and is able to reconstruct equally valid gene regulatory networks.ConclusionGiven that previous versions of ARACNE are extremely resource demanding, either in computational time or in hardware investment, GPU-ARACNE is remarkably valuable for researchers who need to build complex regulatory networks from large expression datasets, but with limited budget on computational resources. In addition, our GPU-centered optimization of adaptive partitioning for Mutual Information (MI) estimation provides lessons that are applicable to other domains.

Inversion of magnetotelluric data using integral equation approach with variable sensitivity domain: Application to EarthScope MT data

We have developed a multi-level parallel magnetotelluric (MT) integral equation based inversion program which uses variable sensitivity domain. The limited sensitivity of the data, which decreases with increasing frequency, is exploited by a receiver sensitivity domain, which also varies with frequency. We assess the effect of inverting principal impedances, full impedance tensor, and full tensor jointly with magnetovariational data (tipper). We first apply this method to several models and then invert the EarthScope MT data. We recover well the prominent features in the area including resistive structure associated with the Juan de Fuca slab subducting beneath the northwestern United States, the conductive zone of partially melted material above the subducting slab at the Cascade volcanic arc, conductive features in the Great Basin and in the area of Yellowstone associated with the hot spot, and resistive areas to the east corresponding to the older and more stable cratons.

Time-dependent current into and through multilevel parallel quantum dots in a photon cavity

We analyze theoretically the charging current into, and the transport current through, a nanoscale two-dimensional electron system with two parallel quantum dots embedded in a short wire placed in a photon cavity. A plunger gate is used to place specific many-body states of the interacting system in the bias window defined by the external leads. We show how the transport phenomena active in the many-level complex central system strongly depend on the gate voltage. We identify a resonant transport through the central system as the two spin components of the one-electron ground state are in the bias window. This resonant transport through the lowest energy electron states seems to a large extent independent of the detuned photon field when judged from the transport current. This could be expected in the small bias regime, but an observation of the occupancy of the states of the system reveals that this picture is not entirely true. The current does not reflect slower photon-active internal transitions bringing the system into the steady state. The number of initially present photons determines when the system reaches the real steady state. With two-electron states in the bias window we observe a more complex situation with intermediate radiative and nonradiative relaxation channels leading to a steady state with a weak nonresonant current caused by inelastic tunneling through the two-electron ground state of the system. The presence of the radiative channels makes this phenomena dependent on the number of photons initially in the cavity.

A GPU Implementation of OLPCA Method in Hybrid Environment

Sophisticated denoising algorithms are used to improve image quality in the Magnetic Resonance Imaging field. Of course, better results are obtained by implementing computationally expensive schemes. In this paper, we consider the Overcomplete Local Principal Component Analysis (OLPCA) method for image denoising and its main issues. More in detail, we investigated the impact of the Singular Value Decomposition on the OLPCA algorithm and its high computational cost. Moreover, we propose a fine-to-coarse parallelization strategy in order to exploit a parallel hybrid architecture and we implement a multilevel parallel software as a smart combination between codes using NVIDIA cuBLAS library for Graphic Processor Units (GPUs) and the standard Message Passing Interface library for cluster programming. Experimental results show improvements in terms of execution time with a promising speed up with respect to the CPU and our old GPU versions.

Corporal Punishment and Child Aggression: Ethnic-Level Family Cohesion as a Moderator.

Ethnicity has been examined as a putative moderator between parents' use of corporal punishment and children's externalizing behaviors. Yet, the reasons for this potential ethnic-level moderator have not been fully examined. The primary objective of this study was to examine whether the effect of corporal punishment on aggression is ethnic-specific using major racial groups inside and outside the U.S. samples and how the mean levels of cohesion in family relationships as found in different ethnic groups moderate the association between mothers' use of corporal punishment and children's aggression. A total of 729 mothers who had children aged 7 to 13 years were sampled from five ethnic groups (i.e., European American, African American, Hispanic American, Korean, and Chinese). Several hypotheses were tested to examine the moderating effect of ethnic-level, family cohesion on the relation of corporal punishment to children's aggression. As expected, the mean level of family cohesion was significantly different across ethnicities. Consistent results across parallel multilevel and fixed effect models showed that high corporal punishment was associated with more aggression in all ethnicities, but there was a significant variation in the association across ethnicities, and the variation was explained by ethnic-level family cohesion. There were weaker associations between corporal punishment and child aggression among ethnic groups with high family cohesion and stronger associations among ethnic groups with low family cohesion. Ethnic/cultural variation in this study emphasizes the importance of understanding family environment of diverse ethnic groups when evaluating the influence of corporal punishment on child behavior in different ethnic/cultural contexts.

Design of Multiple-level Parallel Dynamic Vibration Absorber Based on the Flexible Beam

The amplitude magnification factor dimensionless expression is gained through modeling multi-level parallel dynamic vibration absorber mathematical model combined with the impedance of elastic foundation based on the assumption of elastic foundation. It is found that the influence of elastic foundation on the main system cannot be ignored and the necessity is proposed to optimize the multiple-level parallel system combining the advantage of decaying vibration. The influence of the distance u between the connection and the end on the main system is also analyzed: the less u, the smaller the influence of the beam elastic vibration on the main system. The first frequency will not occur coupling vibration with the main system when u  1m, and u=2m has the most influence. The numerical optimization method is used to design the parameters of 2-grade parallel and 4-grade parallel system. The result shows: the result of rigid connects system is not applicable to the elastic system. It will effectively reduce the influence of the beam elastic vibration on the primary system through optimizing the whole system. At last, the connection dynamic stiffness is researched, it shows that the first natural frequency of beam has no effect on the first three nature frequencies when 40 d K  k , the influence can be neglected when 70 d K  k .

Natural laminar flow shape optimization in transonic regime with competitive Nash game strategy

• NLF shape design in transonic regime is considered as a MCO problem. • A multi-level PNEAs is developed and used to solve NLF optimization efficiently. • Both wave drag and friction drag performances of a NE are greatly improved. • Performance of NE for NLF shape optimization is equivalent to that of Pareto front. The Natural Laminar Flow (NLF) airfoil/wing design optimization is an efficient method which can reduce significantly turbulence skin friction by delaying transition location at high Reynolds numbers. However, the reduction of the friction drag is competitively balanced with the increase of shock wave induced drag in transonic regime. In this paper, a distributed Nash Evolutionary Algorithms (EAs) is presented and extended to multi-level parallel computing, namely multi-level parallel Nash EAs. The proposed improved methodology is used to solve NLF airfoil shape design optimization problem. It turns out that the optimization method developed in this paper can easily capture a Nash Equilibrium (NE) between transition delaying and wave drag increasing. Results of numerical experiments demonstrate that both wave drag and friction drag performances of a NE are greatly improved. Moreover, performance of the NE is equivalent to that of cooperative Pareto-optimum solutions, but it is more efficient in terms of CPU time. The successful application validates efficiency of algorithms in solving complex aerodynamic optimization problem.

A parallel PDE-based numerical algorithm for computing the Optical Flow in hybrid systems

In this paper, we propose a fine-to-coarse parallelization strategy in order to exploit, in a case study, a parallel hybrid architecture. We consider the Optical Flow numerical problem, modelled by partial differential equations, and implement a parallel multilevel software. Our hybrid software solution is a smart combination between codes on Graphic Processor Units (GPUs) and standard scientific parallel computing libraries on a cluster. Numerical experiments, on real satellite image sequences coming from a large dataset in a big data scenario, together with application profiling, highlight good results in terms of performance for the proposed approach.

Exploiting multilevel parallelism on a many-core system for the application of hyperheuristics to a molecular docking problem

The solution of Protein–Ligand Docking Problems can be approached through metaheuristics, and satisfactory metaheuristics can be obtained with hyperheuristics searching in the space of metaheuristics implemented inside a parameterized schema. These hyperheuristics apply several metaheuristics, resulting in high computational costs. To reduce execution times, a shared-memory schema of hyperheuristics is used with four levels of parallelism, two for the hyperheuristic and two for the metaheuristics. The parallel schema is executed in a many-core system in “native mode,” and the four-level parallelism allows us to take full advantage of the massive parallelism offered by this architecture and obtain satisfactory fitness and an important reduction in the execution time.

P-Hint-Hunt: a deep parallelized whole genome DNA methylation detection tool

BackgroundThe increasing studies have been conducted using whole genome DNA methylation detection as one of the most important part of epigenetics research to find the significant relationships among DNA methylation and several typical diseases, such as cancers and diabetes. In many of those studies, mapping the bisulfite treated sequence to the whole genome has been the main method to study DNA cytosine methylation. However, today’s relative tools almost suffer from inaccuracies and time-consuming problems.ResultsIn our study, we designed a new DNA methylation prediction tool (“Hint-Hunt”) to solve the problem. By having an optimal complex alignment computation and Smith-Waterman matrix dynamic programming, Hint-Hunt could analyze and predict the DNA methylation status. But when Hint-Hunt tried to predict DNA methylation status with large-scale dataset, there are still slow speed and low temporal-spatial efficiency problems. In order to solve the problems of Smith-Waterman dynamic programming and low temporal-spatial efficiency, we further design a deep parallelized whole genome DNA methylation detection tool (“P-Hint-Hunt”) on Tianhe-2 (TH-2) supercomputer.ConclusionsTo the best of our knowledge, P-Hint-Hunt is the first parallel DNA methylation detection tool with a high speed-up to process large-scale dataset, and could run both on CPU and Intel Xeon Phi coprocessors. Moreover, we deploy and evaluate Hint-Hunt and P-Hint-Hunt on TH-2 supercomputer in different scales. The experimental results illuminate our tools eliminate the deviation caused by bisulfite treatment in mapping procedure and the multi-level parallel program yields a 48 times speed-up with 64 threads. P-Hint-Hunt gain a deep acceleration on CPU and Intel Xeon Phi heterogeneous platform, which gives full play of the advantages of multi-cores (CPU) and many-cores (Phi).

Novel Multilevel Preconditioners for the Systems Arising from Plane Wave Discretization of Helmholtz Equations with Large Wave Numbers

In this paper we are concerned with fast algorithms for the systems arising from the plane wave discretizations for two-dimensional Helmholtz equations with large wave numbers. We consider the plane wave weighted least squares (PWLS) method and the plane wave discontinuous Galerkin (PWDG) method. The main goal of this paper is to construct multilevel parallel preconditioners for solving the resulting Helmholtz systems. To this end, we first build a multilevel space decomposition for the plane wave discretization space based on a multilevel overlapping domain decomposition method. Then, corresponding to the space decomposition, we construct an additive multilevel preconditioner for the underlying Helmholtz systems. Further, we design both additive and multiplicative multilevel preconditioners with smoothers, which are different from the standard multigrid preconditioners. We apply the proposed multilevel preconditioners with a constant coarsest mesh size to solve two-dimensional Helmholtz systems generated by the PWLS method or PWDG method, and we find that the new preconditioners possess nearly stable convergence; i.e., the iteration counts of the preconditioned iterative methods (PCG or PGMRES) with the preconditioners increase very slowly when the wave number increases (and the fine mesh size decreases).