Sala: LabGrad DCA
Horário: Terça-Feira (03/11/2026), 10:20 - 12:30
Resumo: Este minicurso explora a evolução de duas décadas da programação CUDA, capacitando estudantes, pesquisadores e profissionais a dominarem detalhes arquiteturais da GPU e do compilador. Serão apresentadas estratégias avançadas de otimização, microbenchmarks e análise prática de limites de desempenho para extrair o máximo de desempenho da aplicação.
Sala: LabGrad DCA
Horário: Terça-Feira (03/11/2026), 14:10 - 16:10
Resumo: Este minicurso aborda o uso de novas arquiteturas computacionais em aplicações de processamento de alto desempenho, tomando como estudo de caso os aceleradores Tenstorrent. Essas arquiteturas combinam processadores RISC-V, unidades especializadas de computação, memórias locais e redes de comunicação em chip para implementar um modelo de execução orientado a fluxo de dados. O tema apresenta aderência direta às áreas de Arquitetura de Computadores e Processamento de Alto Desempenho do SSCAD, explorando novas formas de organização, programação e execução de aplicações paralelas.
Sala: Lab DIMAp
Horário: Terça-Feira (03/11/2026), 14:10 - 16:10
Resumo: The heterogeneity of computational resources in modern High Performance Computing (HPC) systems, featuring multicore CPUs, GPGPU accelerators, and multi-node clusters, has made traditional parallel programming paradigms increasingly complex and difficult to port across architectures. The task-based parallel programming paradigm addresses this challenge by transferring responsibilities such as scheduling, data management, and resource mapping to a runtime system, allowing the programmer to focus on defining tasks, their implementations, and data dependencies. This advanced tutorial presents the task-based paradigm and demonstrates how to build, optimize, and analyze parallel applications using the StarPU runtime. Participants will learn to construct Directed Acyclic Graphs (DAGs) from sequential code, define tasks with multiple implementations for heterogeneous resources (CPU and GPU), and leverage StarPU's data flow scheduling model. The tutorial covers practical topics including block-matrix computations, data partitioning, dependency management, and heterogeneous programming with CUDA/cuBLAS. The tutorial also introduces performance analysis techniques for task-based applications, going beyond traditional metrics (makespan, speedup, efficiency) with the StarVZ framework for trace visualization and bottleneck detection. Through hands-on exercises, participants will build complete applications and analyze their performance using execution traces.
Sala: Lab DIMAp
Horário: Quarta-Feira (04/11/2026), 08:00 - 09:50 (extensão opcional até 11:20)
Resumo: A compressão de dados científicos é uma técnica amplamente utilizada para reduzir o consumo de memória, armazenamento e comunicação em aplicações de Computação de Alto Desempenho (HPC). Este minicurso apresenta os fundamentos, principais ferramentas e aplicações da compressão em GPUs, abordando aspectos teóricos e práticos de sua utilização. O tema está diretamente relacionado às áreas de Aplicações Científicas de Computação de Alto Desempenho.
Sala: Lab DIMAp
Horário: Quarta-Feira (04/11/2026), 14:00 - 15:50 (extensão opcional até 18:20)
Resumo: Supercomputers are tools that enable the solution of large-scale models to advance scientific and technological development. However, their processing capacity is not always fully utilized. Due to memory bandwidth and synchronization bottlenecks, practical engineering and scientific applications rarely sustain more than a small fraction, often between 1% and 3%, of the theoretical processing capacity of modern architecture. Traditional software development support tools for supercomputers focus on acceleration in specific configurations, without predicting behavior on other machines. To optimize the use of supercomputers, the Laboratory of Parallel Architectures for Signal Processing at the Federal University of Rio Grande do Norte (UFRN) developed the Parallel Scalability (PaScal) Suite. This innovative set of tools differs by evaluating program scalability rather than focusing solely on performance. Through the PaScal Suite, it is possible to predict how a program behaves across different configurations and machines, optimizing resource utilization and improving overall efficiency. The PaScal Suite is a Brazilian toolset with international recognition and was recently integrated into the VI-HPS (Virtual Institute – High Productivity Supercomputing), highlighting its relevance to the global HPC community. The suite integrates two tools: PaScal Analyzer and PaScal Viewer, simplifying the execution, measurement, and comparison of parallel program runs. It enables the analysis of scalability trends across different processing configurations and workloads, providing visual elements that help identify scalability bottlenecks. This toolset contributes to the development of scalable parallel applications on shared-memory computing nodes. The short course addresses the importance and methods of evaluating scalability in parallel programs, demonstrating how the PaScal Suite can assist developers in profiling and performance analysis, while providing hands-on instruction for identifying critical points and bottlenecks.
