Defesa de Seminário de Andamento de Doutorado (PPGCC/UFSC) - Vanderlei Munhoz Pereira Filho

Defesa de Seminário de Andamento de Doutorado Programa de Pós-Graduação em Ciência da Computação (PPGCC) Universidade Federal de Santa Catarina (UFSC) Título do Trabalho: Scheduling of Task-Based Parallel Applications on Heterogeneous Cloud Computing Environments Autor: Vanderlei Munhoz Pereira Filho Orientadores: Prof. Dr. Márcio Castro, Prof. Dr. Olivier Aumage, Prof. Dr. Laércio Lima Pilla Resumo: High-performance computing applications increasingly run on public cloud infrastructure, where spot instances offer a substantial discount in exchange for the provider’s right to reclaim them at any moment. Exploiting that discount requires applications to survive interruption, and the fault-tolerance mechanisms that make this possible impose a performance cost. This thesis investigates whether task-based parallel applications can run cost-effectively on volatile and heterogeneous cloud infrastructure, and what information the application and the infrastructure management layer must exchange for that to work. The thesis presents three completed contributions. The first is HPC@Cloud, an open-source provider-agnostic toolkit that provisions, monitors and recovers heterogeneous spot clusters directly through provider interfaces, with an explicit cluster state model and a declarative elasticity policy that governs replacement provisioning after an interruption. The second extends StarPU, a scheduling system for task-based parallel applications, with persistent checkpoint-restart and elastic restore, so that a job terminated by an interruption resumes on a cluster whose size and hardware composition may differ from those of the cluster that wrote the checkpoint. The third is a practical characterization of the performance and cost of task-based workloads on public clouds, covering heterogeneous clusters (CPU+GPU) and spot instances. To evaluate HPC@Cloud and the modifications to StarPU, a distributed Full Waveform Inversion (FWI) benchmark was developed. Experimental results on production AWS infrastructure show that the persistent and elastic checkpointing mechanism costs between 1.5 and 5.5 percent of total runtime, and that the cost of a single write is nearly independent of cluster size. Reconstructing application state costs under 300 milliseconds in every topology measured, including the hybrid restore in which state written by ranks running on CPUs is read by a rank with a GPU worker. Recovery latency is therefore governed by the acquisition of replacement capacity in the cloud, and not by the checkpoint mechanism developed here. An interruption reduces the saving obtained with spot instances without eliminating it, since even the most expensive recovery measured cost no more than an uninterrupted run at on-demand infrastructure prices. The remaining doctoral work addresses the fourth contribution, a strategy that exposes infrastructure cost and volatility to the StarPU scheduler, so that task placement answers to monetary objectives in addition to performance ones. Palavras-chave: Cloud computing. High-performance computing. Fault tolerance. Spot instances. Task scheduling.