Julia is a high-level language that supports executing parallel code through various packages. CUDA.jl is prominently used for developing GPU Julia code across a significant number of libraries and programs. In this paper, Juliana, a new tool that translates Julia code utilizing the CUDA.jl package to an abstract multi-backend representation powered by the KernelAbstractions package, is presented. The performance impact of this translation is evaluated using a custom adaptation of the well-established Rodinia benchmark suite to Julia CUDA.jl. To ensure the viability of the tool from a performance perspective, an accurate statistical analysis of the overhead using the BenchmarkTools Julia package is performed, comparing the same benchmark code on the same CUDA device before and after the translation. Additionally, the portability of this approach is demonstrated by running the translated code across multiple backends of KernelAbstractions, allowing the execution of the Rodinia benchmark suite on different GPU vendors such as NVIDIA, Intel, AMD, or Apple.
In this paper we study the power consumption of quantum computing platforms when integrated into high-performance computing (HPC) centers. We analyze the key components of leading quantum computers (superconducting circuit, trapped ion, neutral atom and photonics), and compare their power consumption. We also introduce a quantum energy efficiency metric that combines multiple qubit performance factors and power consumption. Using this metric, we evaluate and compare the energy efficiency of various quantum computers and assess their performance.
The Edge Computing environments enable the development of pervasive applications distributed across extensive geographical areas, overcoming specific issues associated with centralized information processing, such as network bandwidth saturation and the need for large computing infrastructures. This work presents two case studies of deploying applications for environmental monitoring on low-energy and highperformance edge computing devices, employing accelerated artificial intelligence techniques based on GPUs. The first problem entails the classification of various materials within hyperspectral images, while the second problem focuses on identifying floating plastic debris. The applications were validated on a Nvidia Jetson Nano sensor board, demonstrating good accuracy, effectiveness, and energy consumption results.
Abstract. Recent wildfire events in various parts of the world (Canada, the USA, Europe, Australia, etc.) highlight that climate evolution leads to atmospheric patterns that contribute to the development of more destructive fires, revealing complex fire-atmosphere interactions. The HIDALGO2 project addresses several environmental challenges from the perspective of numerical simulation and explores the possibilities and constraints of using European HPC resources. This paper presents some of the strategies followed in the wildfire pilot study for modeling fire spread and smoke emission and dispersion. The WRF-SFIRE model allows for the prediction of atmospheric behavior and its influence on wildfire spread by coupling the factors governing both phenomena. The design of the number of domains and their resolution, as well as the parameterization for downscaling and fire spread simulation, influence performance on EuroHPC installations. Additionally, each HPC installation has particularities for its use. These aspects condition and limit their use for real-time simulations to support operational decision-making. The pre-calculation of simulation ensembles can help overcome this obstacle and also allow for landscape sensitivity analysis. Finally, some strategies for the photorealistic visualization of the results are presented.
A multi-scale modeling approach is employed for the study of the effect of oleic-acid (OA) coverage on the magnetic behaviour of Co ferrite nanoparticles (CFNs), using high performance computing (HPC). Our study is performed in three different length scales: first we perform electronic structure calculations using parallel implementation of the density functional theory (DFT) method to study the magnetic properties of ultra-small OA coated CFNs. Next taking input from the DFT data, we calculate the magnetic characteristics of larger in size OA coated CFNs at an atomic scale performing Monte Carlo simulations. Finally, a mesoscopic modeling approach for interacting assemblies of nanoparticles is employed to reduce further computation time and sources for the study of the magnetic behaviour of CFNs covered with different percentage of oleic-acid, at finite temperature. The results demonstrate that the DFT magnetic moment and magnetic anisotropy of the nanoparticle decrease with the increase of the percentage of the surfactant. However, in the assembly of CFNs the interplay between the exchange and dipolar interparticle interactions results in the increase of the magnetic anisotropy and the decrease of the saturation magnetization as the percentage of OA coverage increases, in agreement with experimental findings. The proposed multi-scale computational approach, implemented in HPC environment, illustrates its ability to handle numerical calculations on complex magnetic interactions of multiple structural components. It can overcome computational limitations to predict optimum parameters for hybrid organic/inorganic nanomaterials for various applications.
The paper describes and analyzes an Average Schwarz Method with spectrally enriched coarse space for a reduced HsiehClough-Tocher (RHCT) finite element discretization of a 4th-order elliptic multiscale problem. The derived symmetric preconditioner is applied and the PCG iterative method is used to solve the preconditioned problem. Suppose the enrichments of the coarse space contain sufficiently many specially constructed eigenfunctions. In that case, the convergence rate of the PCG method is weakly dependent on the ratio of the coarse to fine mesh h/H.
An understanding of the growth of the microbiota in soil is of great importance for the health of farms. This research introduces a new method, PINNs, which captures the intricate relationships between soil inhabitants and newly introduced species. PINNs are more accurate than traditional methods and, when combined with parallel computation, achieve excellent results. This approach promises to improve microbial growth predictions, leading to a deeper understanding of soil health and function.
SYCL programming model does not guarantee performance portability across different architectures. However, the HPC community severely needs platform-independent performance portable applications more than ever. Therefore, the main challenge of SYCL implementers and application developers is to look for direct or indirect solutions in order to improve the portability, performance, and performance portability of SYCL applications. In this paper we study and analyze the impact of the two main SYCL abstractions for data management, i.e., the unified shared memory and buffer-accessor approaches, on the three pillars of performance portability: portability, productivity, and performance. Experiments were carried out on state-of-the-art CPU and GPU platforms in order to shed light on the effect of different SYCL features on performance. The conclusions that emerged from this study show that by avoiding the use of SYCL performance portability inhibitors, it is possible to develop applications with a realistic level of performance portability.
We propose the use of the polar decomposition for the preconditioning of the one-sided block-Jacobi algorithm for the singular value decomposition of a given matrix A. The preconditioner comes from the eigenvalue decomposition of the Hermitian factor H, which is computed by using (partial) Halley's iterations. This approach eliminates the computation of the Gram matrix AT A, which is not numerically reliable for very ill-conditioned matrices A. The iterated matrix in Halley's iterations has a special structure, and three variants for its QR decomposition are proposed and compared. Numerical experiments show, that this new approach is efficient for very ill-conditioned matrices, whereas the Gram matrix can be safely used in other cases.
In this paper, we explore the problem of learning the parameters of Gaussian mixture models using the expectation-maximization (EM) algorithm. We propose a new parallel formulation of the EM algorithm that utilizes a static decomposition and distributes the learning set and the matrix storing posterior probabilities among MPI processes. During both the E-step and M-step of an EM iteration, each MPI process spawns a team of OpenMP threads that process data in blocks sized to fit the last-level cache. The calculations necessary for obtaining weighted Gaussian densities in the E-step and mixture parameters in the M-step are conducted using optimized level 3 BLAS operations. Hierarchical all-reduce operations are employed to compute the sums required in the M-step, initially among the OpenMP threads within each MPI process and subsequently among all the MPI processes in the parallel application. In our computational experiments, we compared this proposed approach with the traditional method that employs level 2 BLAS operations and assessed its strong scaling on 64 nodes of a compute cluster. The results demonstrate that the proposed method is 1.6 to 4.7 times faster than the conventional approach utilizing level 2 BLAS. The parallel efficiency of our approach on 64 nodes ranges from 60
IMRT uses radiation beams with different angles and intensities to target cancerous tissues while protecting healthy organs. Planning methods based on the generalized Equivalent Uniform Dose metric produce plans with excellent tumor coverage, but necessitate the adjustment of many parameters. To address this challenge, a novel approach, PersEUD, has been proposed for the automated tuning of these parameters. This is achieved by combining solutions from a Gradient Descent algorithm with an evolutionary optimization method to explore the parameter space efficiently. Previous research has demonstrated the effectiveness of this approach in meeting clinical constraints. However, its high computational demands hinder its integration into clinical practice. The goal of this study is to accelerate the optimization processes by distributing the evaluations in the nodes of modern multicore clusters. At the node level, these evaluations can be efficiently computed with the combination of parallelization and batching strategies. As a consequence, the efficiency of the evaluations depends on the node's load, and the distribution of evaluations among the nodes must account for this dependence. In this study, we propose an approach to integrate an efficient scheduling of evaluations on heterogenous multi-core nodes in PersEUD. The proposal has been extensively tested on eight clusters, with nodes of three different micro-architectures. The test data set consisted of three head and neck patients treated with IMRT using nine beams. The results indicate that exploiting the cluster appropriately leads to a substantial acceleration of the computation involved in the planning based on PersEUD. This result facilitates the practical implementation of PersEUD in clinical settings.
Secure Multiparty Computation (SMC) facilitates secure collaboration among multiple parties while safeguarding the privacy of their confidential data. This paper introduces a two-party quantum SMC protocol designed for evaluating binary Boolean functions using single qubits. Complexity analyses demonstrate a reduction of 66.7% in required quantum resources, achieved by utilizing single qubits instead of multi-particle entangled states. However, the quantum communication cost has increased by 40% due to the amplified exchange of qubits among participants. Furthermore, we bolster security by performing additional quantum operations along the y-axis of the Bloch sphere, effectively hiding the output from potential adversaries. We design the corresponding quantum circuit and implement the proposed protocol on the IBM Qiskit platform, yielding reliable outcomes.
Various macroscopic and microscopic road traffic models allow traffic flow analysis. However, it should be emphasised that standard traffic flow models do not include drivers' behaviour. Thus, we propose a multi-agent microscopic model for analysing the roundabout traffic flow considering the various types of agents. Agents have parameters characterising their style of acceleration and braking, as well as the distance to the vehicle in front. The simulation studies show that these parameters are crucial in roundabout road traffic analysis. To accurately reflect the acquired dimensions of the cars, a small-cell cellular automaton (CA) was used, where one car is represented as a set of CA cells.
In recent years, a variety of new frameworks streamlining the process of agent-based modeling has emerged. These frameworks serve different purposes and each offers a unique set of features. In this practical comparative study, we evaluate the performance of various ABM frameworks through a series of benchmark simulations. By comparing the distinct functionalities offered by these tools, we aim to assist in the selection of an appropriate ABM toolkit for developing system models. This review presents a concise overview of seven popular agent-based modeling tools aiming to inspire further exploration and investigation into this subject.
This paper delves into the application of Uncertainty Quantification (UQ) and Sensitivity Analysis (SA) to address complex, multiscale Global Challenges. Using a Renewable Energy Sources case study, we demonstrate various approaches to incorporate UQ and SA. UQ helps mitigate uncertainties in models and input data, leading to more reliable results. SA identifies the significant influence of specific input parameters on model outputs, aiding in resource allocation and problem-solving. By reducing the number of required parameters, SA can optimize computational resources and accelerate time-to-solution. Additionally, we showcase how UQ and SA can directly contribute to addressing Global Challenges. The paper concludes by discussing the multiscale Uncertainty Quantification and Sensitivity Analysis platform (mUQSA) and its underlying tools, which streamline the implementation of UQ and SA for Global Challenges.
Scientific computing centers or private (in-house) cloud data centers do not rely on the standard pay-as-you-go business model which is common in commercial clouds to allocate resources. Instead, the system is typically shared by a set of selected users, and the administrator's job is to ensure that resources are shared fairly given the existing policies of that organization. One common approach, especially in batch systems, is to deploy a fairshare-based prioritization in the scheduler, where a prioritization mechanism balances resource consumption so that individual users get the right shares of resources over time. In this paper, we present a tool developed to simulate the fairshare setting in a batch system. Using a set of experiments, we demonstrate the utility of this tool in tuning fairshare settings in a standard HPC/HTC scheduler and present the impact of often-overlooked additional options for modifying the basic fairshare settings. All the findings in this paper are based on our real-world experience of running and optimizing a distributed national computing infrastructure in the Czech Republic.
Random number generation is crucial for secure communications and simulations. While classical generators have inherent limitations due to determinism, quantum random number generators offer a safer alternative. However, most generators depend on assumptions about their hardware, posing a risk if compromised. This paper presents a semi-device-independent protocol, reducing reliance on the generators' components. The scheme is based on the prepare-and-measure scenario and operates under an easy-to-verify assumption related to the laser's mean- photon number, employing a CMOS sensor as the practical detector. Using a correlation test of preparation and measurement, we validate the generated randomness even with untrusted devices.
Graph Neural Networks (GNNs) have emerged as powerful tools for learning on graph-structured data, demonstrating state-of-theart performance in various applications such as social network analysis, biological network modeling, and recommendation systems. However, the computational complexity of GNNs poses significant challenges for scalability, particularly with large-scale graphs. Parallelism in GNNs addresses this issue by distributing computation across multiple processors, using techniques such as data parallelism and model parallelism. Data parallelism involves partitioning the graph data across different processors, while model parallelism splits the neural network's layers or operations. These parallelization strategies, along with optimizations such as asynchronous updates and efficient communication protocols, enable GNNs to handle larger graphs and improve training efficiency. This work explores the key computational kernels and looks for those where parallelism could significantly enhance the scalability and performance of GNNs, highlighting the algebraic aspects of each one. This is the first step to better compare recent advances and their implications for future research.
The field of laser-ion acceleration faces significant challenges in handling high-dimensional, computationally intensive problems, often constrained by budgets and available computational power. Reliably achieving high ion energies with current laser technologies through ultrahigh intensity pulses in near-critical to overdense plasmas remains difficult, necessitating detailed, costly simulations to explore various acceleration mechanisms and optimize outcomes. Due to the high computational cost of these simulations, there is a need for precise, scalable, and efficient adaptive sampling methods that balance exploring new mechanisms while exploiting known parameter dependencies and that can operate in high-performance computing environments. In this work, we propose and investigate a scalable adaptive sampling approach that intrinsically supports parallel processing. We apply this method to data obtained from particle-in-cell simulations using multilayer perceptrons (MLPs), chosen for their flexibility in modeling complex dependencies. We benchmark our results against Bayesian optimization, and we highlight the limitations of MLPs when accounting for the intrinsic uncertainty present in the acceleration process.
With the inclusion of Aurora in the TOP500 list in November 2023 three different GPU (Graphics Processing Unit) vendors are represented in the top 10 of the list. Each vendor has its own preferred model for programming GPUs. For this paper we implemented the Nbody problem using portable programming frameworks and the vendors' preferred APIs. We show how the performance of the portable solutions compares to the performance of the native solution on each hardware both in absolute numbers and as raction of the achievable peak performance.
