Program

Dr. Vanesa M. Tennant Williams, Moderator

CARLA 2025 Welcome Message
Dr. Kevin A. Brown
General Co-Chair
CARLA 2025 

Remarks
Dr. Charah T. Watson
Executive Director
Scientific Research Council

Prof. Tannecia Stephenson
Deputy Dean, Faculty of Science and technology
Co-Director of the Climate Studies group, Mona (CSGM)
University of the West Indies, Mona

Mrs. Anika C. D. Shuttleworth
Chief Information Officer
JAMICTA - ICT Authority

Dr. Carlos Jaime Barrios Hernandez
General Chair
SCALAC

Abstract. NASA’s missions span human and robotic space exploration, ground-breaking aeronautics research, and Earth and space sciences. This talk will provide a broad overview of numerous HPC and related technologies that NASA develops, adapts, and implements for its wide spectrum of programs and projects ranging from Earth to deep space.

Biography. Dr. Rupak Biswas is currently the Director of Exploration Technology at NASA Ames Research Center, Moffett Field, Calif., and has held this Senior Executive Service (SES) position since January 2016. In this role, he is in charge of planning, directing, and coordinating the technology development and operational activities of the organization that comprises of advanced supercomputing, human systems integration, intelligent systems, and entry systems technology. The directorate consists of approximately 700 employees with an annual budget of $160 million, and includes two of NASA’s critical and consolidated infrastructures: arcjet testing facility and supercomputing facility. He is also the Manager of the High End Computing Capability Project that provides a full range of advanced computational resources and services to numerous NASA programs. In addition, he leads the emerging quantum computing effort for NASA. Dr. Biswas received his Ph.D. in Computer Science from Rensselaer in 1991, and has been at NASA ever since. During this time, he has received several agency awards, including the Exceptional Achievement Medal and the Outstanding Leadership Medal. He is an internationally recognized expert in high performance computing and has published more than 150 technical papers, received many Best Paper awards, edited several journal special issues, and given numerous lectures around the world.

What you "REALLY" need to know about #AI (from AI to AgenticAI)
by Francisco Aguirre (Dell Technologies) & Pedro Mario Cruz e Silva  (NVIDIA)

Francisco Aguirre,
LATAM NVIDIA Solutions Senior Principal, Dell Technologies

Francisco Aguirre is an experienced technology leader with more than 30 years in the IT industry, specializing in Artificial Intelligence, High-Performance Computing, and emerging technologies. He currently leads NVIDIA solutions for Dell Technologies in Latin America, helping organizations harness the power of accelerated computing to drive innovation and competitive advantage.

Throughout his career, Francisco has advised clients across key industries—including finance, telecommunications, retail, airlines, and education—on how to adopt and scale transformative technologies.

His expertise spans from data analytics, business intelligence, and big data, to modern AI deployments leveraging NVIDIA platforms, GPU-based architectures, and hybrid cloud strategies.

Francisco is recognized as a trusted advisor, speaker, and thought leader. He has delivered keynotes and technical sessions at major industry events such as Dell Technologies World, Dell Technologies Forum, and Mexico Business Forum.

His presentations focus on making cutting-edge concepts like Generative AI, Retrieval-Augmented Generation, Agentic AI, and Quantum Computing accessible to both business and technical audiences.

He holds a degree in Systems and Computer Science Engineering from La Salle University and a master’s degree in Customer Relationship Management from Duke University. Passionate about innovation, Francisco continues to drive conversations at the intersection of technology, business, and culture.

Pedro Mário Cruz e Silva,
Senior Solutions Architect | NVIDIA Latin America

Pedro Mário Cruz e Silva did his BSc (1995), and MSc (1998) at Federal University of Pernambuco (UFPE), he also did his DSc in 2004 at PUC-Rio. He created the Computational Geophysics Group atPUC-Rio were worked for 15 years as Manager, during this period was responsible for several Software Development and R&D projects for Geophysics with strong focus on innovation. He also finished an MBA in 2015 at Getúlio Vargas Foundation (FGV/RJ). Currently is Senior Solutions Architect for Higher-Education and Research for the Latin America Region.

Served in the Grand Jamaica Suite pre-function area

11:40 AM - 1:00 PM
Chair: Esteban Mocskos

1. Machine Learning for Predicting Job States and CPU Power on a Supercomputer
2. Performance and Energy Consumption Prediction of Scientific Workflows using Machine Learning
3. Profiling a task-based molecular dynamics application with a data science approach
4. Investigating the Impact of DVFS on the Energy Efficiency of AI Workloads on GPUs

Authors. Dylan Benavides Castillo (Costa Rica National High Tecnology Center - CENAT); Fabricio Quirós Corella (Costa Rica National High Tecnology Center - CENAT); Esteban Meneses (Costa Rica National High Technology Center - CENAT)

Abstract. Efficient resource management in high-performance computing (HPC) is essential for optimizing costs, reducing energy consumption, and improving system productivity. However, job variability and failures introduce uncertainties that complicate scheduling and resource allocation. Accurately predicting job failures and estimating energy consumption can enhance planning and operational efficiency. This study analyzes data from the Simple Linux Utility for Resource Management (SLURM) on the Kabré supercomputer at Costa Rica’s National High Technology Center (CeNAT). After selecting and preprocessing relevant variables, a dataset was created to train a two-stage machine learning model comprising a binary classifier and a regression model. Using 10-fold cross-validation, multiple models were evaluated, with Random Forest emerging as the best performer in both stages. The classification model was assessed using the confusion matrix and ROC curve, while the regression model was evaluated through residual analysis and metrics such as Root Mean Square Error (RMSE) and the Coefficient of Determination (R^2). This approach can support users and administrators by improving job scheduling decisions and reducing energy waste in HPC systems. 

Authors. Felipe Barbosa (Federal University of Pará - UFPA); Josivaldo de Souza Araújo (Federal University of Pará - UFPA); Marcos Amarís (Federal University of Pará - UFPA); Erick Damasceno (Federal University of Pará - UFPA); Fellipe Queiroz (Federal University of Pará - UFPA); Josiany Brito Guimarães (Federal University of Pará - UFPA); Daniel Cordeiro (University of São Paulo - USP)

Abstract. In High Performance Computing (HPC), large-scale scientific workflows are essential for modern discoveries but lead to significant energy consumption. This work explores predictive models to estimate both energy consumption and performance, to support sustainable computing in HPC environments. We used WfCommons to generate workflows, Wrench to simulate the supercomputing environments, and Scikit-learn to implement machine learning algorithms. Regression models, including ensemble techniques, were developed and evaluated using widely adopted scientific workflows such as BLAST, Montage, and Epigenomics. For training, features included IO time (in seconds) and the amount of bytes read and written. In energy consumption prediction, the Gradient Boosting Regressor (GBR) achieved high R^2 scores, such as 0.8556 for Epigenomics and 0.7143 for BLAST. For performance prediction, GBR also showed superior accuracy, with MAE and MAPE as low as 0.0257 and 0.0068, respectively, in the BLAST workflow. These results confirm the effectiveness of ensemble models in energy efficiency and performance, contributing to sustainable scientific computing.

Authors. Christian Asch (Costa Rica National High Technology Center - CENAT); Lucas Mello Schnorr (Federal University of Rio Grande do Sul - UFRGS); Esteban Meneses (Costa Rica National High Technology Center - CENAT)

Abstract. Charm++ is a parallel programming framework based on task-driven execution and global object references. It has been used successfully in various high-performance computing (HPC) applications, including the molecular dynamics simulator NAMD. While Charm++ includes built-in support for performance tracing and visualization through its Projections tool, the existing system offers limited extensibility and has no support for modern data science workflows. This work presents a new visualization and analysis pipeline for Charm++ trace data that emphasizes modularity, openness, and composability. Our toolchain leverages standard scripting languages and data formats - producing output in CSV and Parquet formats - to facilitate integration with data analysis ecosystems. We demonstrate the effectiveness of this approach using LeanMD, a proxy application derived from NAMD, and highlight specific types of custom visualizations that are difficult to achieve with Projections. Our system enables custom visualizations and streamlined analysis of chare-level execution behavior, offering researchers and tool developers improved capabilities for understanding program performance and identifying load imbalance. We discuss the architecture of our tool, its application to real-world traces, and potential extensions for other task-based frameworks.

Authors. Arthur Lorenzon (Federal University of Rio Grande do Sul - UFRGS); Thiago Goncalves (Federal University of Rio Grande do Sul - UFRGS)

Abstract. The increasing scale of artificial intelligence (AI) models has led to unsustainable energy consumption in GPU-based systems, creating an important need for more efficient computing strategies. While Dynamic Voltage and Frequency Scaling (DVFS) can be an efficient technique, the combined impact of GPU core and memory frequencies is often not systematically evaluated. This paper presents a comprehensive analysis of how tuning both GPU core and memory frequencies affects performance, energy consumption, and the Energy-Delay Product (EDP) for AI workloads. We evaluate seven benchmarks with diverse computational demands on an AMD Radeon RX 7700XT GPU across 12 distinct frequency configurations. Our results reveal that memory-bound applications can benefit from the increase of memory frequency, reducing execution time by up to 80.7%, and compute-bound applications benefited more from higher core frequencies. In addition, by selecting appropriate operating frequencies, it is possible to reduce the EDP by up to 98.3% comparing to the worst-case scenario, highlighting the possible energy efficiency gains of choosing a balanced configuration that optimizes the energy-performance trade-off.

Ricardo Baeza-Yates
KTH Royal Institute of Technology, Sweden
Universitat Pompeu Fabra, Barcelona
Universidad de Chile

Abstract. Machine learning (ML), particularly deep learning, is being used everywhere. However, not always is used well, ethically and scientifically. In this talk we first do a deep dive in the limitations of supervised ML and data, its key component. We cover small data, datification, bias, predictive optimization issues, evaluating success instead of harm, and pseudoscience, among other problems. The second part is about our own limitations using ML, including different types of human incompetence: cognitive biases, unethical applications, no administrative competence, misinformation, and the impact on mental health. In the final part we discuss regulation on the use of AI and responsible AI principles, that can mitigate the problems outlined above.

Biography. Ricardo Baeza-Yates is a a part-time WASP Professor at KTH Royal Institute of Technology in Stockholm, as well as part-time professor at the departments of Engineering of Universitat Pompeu Fabra in Barcelona and Computer Science of University of Chile in Santiago. Before, he was VP of Research at Yahoo Labs, based in Barcelona, Spain, and later in Sunnyvale, California, from 2006 to 2016. He is co-author of the best-seller Modern Information Retrieval textbook published by Addison-Wesley in 1999 and 2011 (2nd ed), that won the ASIST 2012 Book of the Year award. In 2009 he was named ACM Fellow and in 2011 IEEE Fellow. He has won national scientific awards in Chile (2024) and Spain (2018), among other accolades and distinctions. He obtained a Ph.D. in CS from the University of Waterloo, Canada, and his areas of expertise are responsible AI, web search and data mining plus data science and algorithms in general.

Moderator: Addison Snell (Intersect360)

4:35 PM - 5:55 PM
Chair: Kyle Felker, Tadeu Gomes

1. Evaluating Malleable Job Scheduling in HPC Clusters using Real-World Workloads
2. Optimizing the Energy-Efficiency of QMCPACK on Aurora Supercomputer via GPU Sharing
3. Good Sustainability Practices for Data Center: A Systematic Literature Review
4. Parallel/distributed computing for optimizing investment planning in electricity markets

Authors. Patrick Zojer (University of Kassel); Jonas Posner (University of Kassel); Taylan Özden (Technical University of Darmstadt)

Abstract. Optimizing resource utilization in high-performance computing (HPC) clusters is essential for maximizing both system efficiency and user satisfaction. However, traditional rigid job scheduling often results in underutilized resources and increased job waiting times. This work evaluates the benefits of resource elasticity, where the job scheduler dynamically adjusts the resource allocation of malleable jobs at runtime. Using real workload traces from the Cori, Eagle, and Theta supercomputers, we simulate varying proportions (0–100%) of malleable jobs with the ElastiSim software. We evaluate five job scheduling strategies, including a novel one that maintains malleable jobs at their preferred resource allocation when possible. Results show that, compared to fully rigid workloads, malleable jobs yield significant improvements across all key metrics. Considering the best-performing scheduling strategy for each supercomputer, job turnaround times decrease by 37–67%, job makespan by 16–65%, job wait times by 73–99%, and node utilization improves by 5–52%. Although improvements vary, gains remain substantial even at 20% malleable jobs. This work highlights important correlations between workload characteristics (e.g., job runtimes and node requirements), malleability proportions, and scheduling strategies. These findings confirm the potential of malleability to address inefficiencies in current HPC practices and demonstrate that even limited adoption can provide substantial advantages, encouraging its integration into HPC resource management.

Authors. Matheus Costa (Federal University of Rio Grande do Sul - UFRGS); Philippe Navaux (Federal University of Rio Grande do Sul - UFRGS); Silvio Rizzi (Argonne National Laboratory); Arthur Lorenzon (Federal University of Rio Grande do Sul - UFRGS)

Abstract. As high-performance computing advances toward the exascale, energy efficiency has become a primary concern. However, complex scientific applications often underutilize GPU hardware, such as GPUs, leaving valuable resources idle. In this scenario, this paper investigates GPU sharing as a strategy to improve resource utilization and energy efficiency for the QMCPACK application on the Aurora supercomputer. We evaluate Intel's Multiple Compute Command Streamers (CCS) technology across a scale of 1 to 16 nodes on Aurora, comparing a baseline single-rank-per-GPU-tile configuration (1-CCS) against multi-rank setups with two (2-CCS) and four (4-CCS) ranks per tile. Our results demonstrate that GPU sharing via the 2-CCS mode yields significant performance improvements regarding the Figure of Merit (FOM) by an average of 17% and enhances energy efficiency (FOMe) by 32% compared to the baseline configuration that uses 1-CCS only. We also show that the 4-CCS configuration suffers from increased MPI communication overhead and lower vectorization efficiency, being worse for performance and energy than running with only 1-CCS. On the other hand, 2-CCS achieves a better balance between higher compute unit engagement and reduced memory system stalls. We also show that exploiting GPU sharing can be an effective strategy for boosting both throughput and sustainability on modern HPC systems, but over-partitioning can introduce overheads that cancel out the benefits.

Authors. Josiany Brito Guimarães (Federal University of Pará - UFPA); Felipe Barbosa (Federal University of Pará - UFPA); Erick Damasceno (Federal University of Pará - UFPA); Fellipe Queiroz (Federal University of Pará - UFPA); Marcos Amarís (Federal University of Pará - UFPA)

Abstract. The search for sustainable practices and energy efficiency in data centers has become a top priority in today’s world, driven by the rapid growth in internet use and online traffic. This highlights the urgent need for research that focuses on practical strategies to reduce the environmental impact of these vital facilities. This article reviews 10 relevant studies that examine the primary methods used to enhance energy efficiency in data centers, including emerging technologies, resource management techniques, and sustainable policies. Additionally, the paper discusses the environmental impacts of data centers, including high energy consumption, greenhouse gas emissions, and intensive use of natural resources. The findings indicate that implementing good sustainability practices can bring several benefits, such as lowering operating costs and reducing the carbon footprint, while maintaining service quality.

Authors. Santiago Freire (Universidad de la República); Sergio Nesmachnow (Universidad de la República); Pedro Moreno (Universidad Autonoma del Estado de Morelos - UAEM)

Abstract. This article presents a parallel and distributed computing strategy for optimizing investment planning in electricity markets. An implementation is developed for a stochastic dynamic programming model used by the Uruguayan electrical company to determine investment strategies in generation assets. The sequential execution of thousands of solver instances results in high computing times, rendering the process inefficient for large-scale scenarios. A distributed architecture based on Message Passing Interface is proposed to enable parallel execution across multiple computing resources. The solution includes modular components organized in a layered architecture, load balancing and fault recovery features. Performance results indicate that the parallel implementation allows addressing complex scenarios with high computing demands. Significant reductions in execution time were obtained, with high speedup values (up to 98x) and efficiency up to 0.86 for a full-scale scenario executed on distributed nodes of a high-performance computing cluster.