Scaling Al Science with ACCESS Pegasus
Автор: SoCal-AIRE
Загружено: 2026-02-06
Просмотров: 5
Описание:
Mats Rynge, University of Southern California
Presented at the NSF ACCESS Regional AI Workshop - SoCal Edition workshop, January 22, 2026.
Abstract: Pegasus WMS (Workflow Management System) streamlines the execution of complex AI and machine learning workloads by automating the end-to-end pipeline from data ingestion to model evaluation. Through ACCESS Pegasus, researchers can utilize a hosted workflow environment that simplifies the orchestration of jobs across distributed national cyberinfrastructure. This platform allows users to leverage pre-configured Jupyter Notebook examples and the Pegasus Python API to design reproducible AI workflows.
To optimize the use of specialized hardware, Pegasus utilizes glideins (pilot jobs) to provide a unified overlay over GPU resources. This abstraction layer allows the workflow manager to treat diverse, distributed compute nodes as a single, coherent pool of resources. By deploying these pilot jobs, Pegasus can dynamically provision and manage high-performance GPU environments, enabling AI workloads to scale across multiple clusters while maintaining consistent performance and reducing the overhead typically associated with manual resource allocation.
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
