Defesa de Exame de Qualificação de Mestrado (PPGCC/UFSC) - Nicolas Vanz

Master's Degree Qualifying Examination Defense Postgraduate Program in Computer Science (PPGCC) Federal University of Santa Catarina (UFSC) Thesis Title: Runtime Support for GPU-Enabled Execution in StarPU Distributed Server-Client Task System Advisor: Prof. Dr. Márcio Castro Abstract: High-performance computing applications increasingly require heterogeneous and distributed execution to meet solution time constraints. In this context, runtimes must coordinate computation, data movement, and synchronization between Central Processing Units (CPUs), Graphics Processing Units (GPUs), and distributed memory nodes, preserving correctness and efficiency. StarPU, a task-based runtime for heterogeneous platforms, offers two main distributed modes: full Message Passing Interface (MPI) and server-client. However, in the server-client path, remote GPU execution was not available as an integrated, lane-type-sensitive flow (execution slot type, CPU, or Compute Unified Device Architecture (CUDA) exposed by each worker node). This work addresses this gap through runtime-level modifications to enable GPU support in the StarPU server-client. The implemented extensions include explicit GPU discovery of worker nodes, worker metadata for CPU/CUDA lane characterization, implementation resolution and lane-sensitive scheduling eligibility, protocol extensions for remote GPU lookup and dispatch, and transfer and completion semantics for source-sink and sink-sink communication. The evaluation was conducted on AWS infrastructure (g6.2xlarge, NVIDIA L4 GPU), with StarPU 1.4.7, OpenMPI 5.0.7, and CUDA 12.9, using an N-body application with direct formulation (O(n2)) as a validation payload. The results show that the proposed approach enables distributed heterogeneous execution in server-client mode and maintains scalability in computationally intensive scenarios; however, it presents additional orchestration overhead and higher execution costs compared to the full MPI model. Keywords: Task-based runtime. Distributed heterogeneous computing. Remote GPU execution. StarPU. MPI.