There are several important science and engineering problems that require the coordinated execution of multiple high-performance simulations. Some common scenarios include but are not limited to, "an ensemble of tasks", "loosely-coupled simulations of tightly-coupled simulations" or "multi-component multi-physics simulations". However, historically supercomputing centers, have supported and priortised the execution of single "jobs" on supercomputers. Not suprisingly, the tools and capabilities to support coordinated multiple simulations are limited.
A promising way to overcome this common limitation is the use of a Pilot-Job --- which can be defined as a container or placeholder job to provide multi-level scheduling via an application-level scheduling overlay over the system scheduler. We discuss both the theory and practise of Pilot-Jobs: Specifically, we introduce the P* Model of Pilot-Jobs and present "BigJob" as a SAGA-based extensible, interopable and scalable implementation of the P* Model. We then discuss several science problems that have/are using BigJob to execute multiple simulations at unprecedented scales on a range of supercomputers and distributed supercomputing infrastructure such as XSEDE.
This talk was given as part of our MSc in HPC's 'HPC Ecosystem' course.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More