Samuel Xavier-De-Souza

Least‐squares reverse time migration (LSRTM) is a powerful seismic imaging technique that estimates reflectivity models by iteratively minimizing the discrepancy between observed and modelled data, using the first‐order scattering approximation of the wave equation. However, because LSRTM relies on Gaussian statistical assumptions, it is susceptible to non‐Gaussian noise, which is frequently encountered in seismic datasets. To overcome this limitation, we introduce a robust generalization of LSRTM based on ‐generalized statistics, termed ‐LSRTM. We formulate a generalized objective function based on the maximum‐likelihood principle that enhances robustness to erratic data while maintaining the computational efficiency of adjoint‐based gradient optimization. Numerical experiments using a complex deep‐water pre‐salt model indicate that the ‐LSRTM framework yields higher‐resolution seismic images with improved geological continuity, especially in deeper subsurface regions, even when Gaussian noise is present. In scenarios with non‐Gaussian noise, ‐LSRTM substantially outperforms conventional LSRTM by providing more accurate reflectivity estimates and achieving faster convergence. These findings demonstrate the practical value of ‐LSRTM for industrial‐scale seismic imaging in environments characterized by challenging noise and complex subsurface structures.

The efficient allocation of resources in high-performance computing (HPC) depends on accurate analysis of application behavior. However, these scalability analyses are often computationally expensive and time-consuming. Although there are tools that automate this process, such as PaScal Suite, the cost of execution is still a challenge. To mitigate these costs, prediction strategies have proven effective, notably the use of neural networks as regression models. The obstacle, however, lies in choosing the data to train these models: the process can be slow and expensive, as depending on the order of the data, it may require that nearly all of the application’s settings be configured in PaScal Suite. This work proposes a strategy based on Reinforcement Learning (RL) to optimize the choice of samples in performance modeling. The main goal is to minimize the number of executions required to train a regression model, while ensuring high accuracy. The methodology employs a Deep Q-Learning agent that iteratively selects the ideal configurations (number of cores and problem size). These selected data are then used to train a neural network capable of predicting the execution time for each configuration, always aiming to maximize information gain and minimize time spent. Experimental analyses were performed to evaluate this approach in comparison to random and heuristic methods. Results demonstrate that the strategy reduced the total exhaustive execution time by approximately 71% and outperformed the random search baseline by nearly 40%, all while maintaining an error rate of less than 10%. This validates the feasibility of using RL to drastically reduce time costs in scalability prediction without compromising accuracy.

Blind deconvolution is a well-established technique in seismic data processing, used to estimate subsurface reflectivity from recorded signals. However, traditional methods often rely on the ℓ₂-norm and assume Gaussian noise, which can limit their performance in real seismic environments, where impulsive and non-Gaussian noise may occur. In order to address this limitation, this work proposes a robust blind deconvolution approach based on a correntropy cost function. By replacing the traditional ℓ₂-norm with a correntropy criterion, the method accounts for higher-order statistical information and becomes robust to outliers and non-Gaussian noise. The algorithm alternates between estimating the reflectivity signal and the source wavelet, based on the Alternating Minimization Sparse Multichannel Blind Deconvolution (AM-SMBD) framework. The method is evaluated on synthetic data under noiseless, Gaussian, and impulsive noise conditions, as well as on real seismic records. Results show that the proposed approach performs similarly to the ℓ₂-based method under Gaussian noise and outperforms it in environments with α-stable non-Gaussian noise. It also was able to better preserve reflectors and attenuate noise in real seismic data.

High-performance computing systems have become increasingly large, dynamic, and complex, making building software that uses full-system resources difficult. While established profiling tools are excellent at identifying performance hotspots and bottlenecks in single runs, they are generally unsuitable for evaluating an application’s overall scalability trends across multiple configurations. This work presents PaScal Suite, a set of tools for automatically measuring and collecting data from the execution of a parallel application with different input arrangements and numbers of processing units in a shared-memory system. In addition to manual instrumentation, the analyzer tool automatically instruments code regions for investigation. It employs trace-based measurement techniques to identify boundaries and collect the required metrics to calculate efficiency values for the regions of interest in each execution condition, with minimal overhead. The viewer tool uses color maps to show variations in efficiency values, highlighting possible hierarchical scalability trends per region. We show that the analyzer tool is minimally intrusive and demonstrate that the suite supports the development and optimization of scalable parallel applications. By combining this automated multi-run workflow with a visual paradigm designed to reveal scalability trends, PaScal Suite offers a highly productive approach to analysis and optimization.

Traditional profilers struggle with scalability analysis. The PaScal Suite’s Viewer tool addresses this, but its server-side architecture creates performance bottlenecks with large datasets. This paper proposes a client-side Single-Page Application (SPA) architecture that transfers all data processing to the browser, enabling fluid, instantaneous analysis. We also introduce a novel hierarchical visualization for comparing code regions, which simplifies diagnostics. The result is an agile platform that enhances the identification of scalability bottlenecks.

With the physical limits of modern processors, parallel programming has become essential to fully harness the potential of current architectures. However, challenges such as race conditions and deadlocks increase the complexity of parallel development. Bend, a functional language with native parallelism, proposes an immutability-based approach to simplify this process. This article analyzes how immutability and native parallelism in Bend eliminate shared memory conflicts, enabling safe and effortless parallelism without explicit synchronization. Through a case study involving tree summation, we compare Bend with a mutable approach in C.

A modelagem sísmica é amplamente usada na exploração de petróleo e gás para gerar imagens da subsuperfície. Apesar de eficaz, exige alto poder computacional. Este trabalho propõe um método mais eficiente, reduzindo tempo de execução e uso de recursos. A técnica tradicional resolve a equação da onda em todo o domínio, o que gera alto custo. A proposta limita a simulação a um volume esférico ao redor da fonte sísmica, definido pelo raio de propagação. O algoritmo foi desenvolvido com essa geometria e testado em diferentes cenários. Os resultados mostram redução significativa no tempo de execução em comparação ao método tradicional.

This study proposes a method to reduce computational cost in time-domain wave modeling. By applying spherical geometric modeling, the wave equation is solved only in dynamically defined regions. The approach was evaluated using mean squared error (MSE) on 3D acoustic models with varying source positions and propagation times. Results showed up to 25% runtime reduction, highlighting the method’s potential for seismic imaging and reservoir analysis.

The Parallel Code Phase Search (PCPS) is a key algorithm in Global Navigation Satellite System (GNSS) receivers for detecting satellites and estimating signal parameters. This work analyzes its scalability in the GNSS-SDR framework, testing varying core counts (1–128) and problem sizes (2–32 million samples). Results show diminishing returns beyond 8–16 cores, with efficiency dropping below 10% at more than 32 cores, regardless of workload. The study identifies parallelization bottlenecks, suggesting optimizations are needed for real-time, high-channel GNSS Software-Defined Receiver (GNSS-SDR) applications on modern multi-core systems.

O software Mamute implementa uma solução numérica para a equação da onda, problema com alto custo computacional na sísmica. Este algoritmo rodava originalmente apenas em CPU com multithreading, e sua adaptação em GPU via OpenMP será comentada neste trabalho. Ademais, comparou-se as performances dos compiladores GCC e Clang, com este último sendo sempre superior em GPU.