Presenter
Abstract
Hybrid MPI programming is a popular approach to program supercomputers today. It allows programmers to reuse existing MPI codes, gives detailed control over data placement and synchronization on the process level, and creates enough opportunity for thread- and device-level programming models to keep resources busy. However, strict ordering of work, manual synchronization, and a fork-join parallelism originating historically from threading control make programming difficult. Consequently, good hardware utilization and performance are hard to achieve.
This talk presents OmpSs-2 and its three productivity features for MPI hybrid programming: The inclusion of MPI communication in the application data-flow, hierarchical tasking to reduce runtime hotspots of master-slave programming models, and dynamic load balancing. We believe that these features help to establish an orthogonal relationship between (MPI-) communication and concurrency, thus simplifying hybrid code development.
About the Seminar:
Available on Blue Jeans: https://bluejeans.com/103029375?src=calendarLink .