Barcelona Super Center tutorial - Session 1

Опубликовано 17 августа 2016, 20:47

The purpose of the seminar is to describe the StarSs programming model, the status of the different implementations and why we consider that it provides a global solution to the challenges raised by future processor and system architectures. StarSs is a task-based programming model where the user annotates with directives an otherwise standard sequential code that is then automatically parallelized by the run time infrastructure. The underlying execution model is a medium-grain data-flow that makes the application progress to dynamically adapt to the huge variability and heterogeneity that we can expect from systems in the future. StarSs also addresses issues related to memory management by separating the concerns of the programmer who specifies the flow of data from those of the run time that optimizes the storage management locality and coherence. Finally, StarSs addresses, either alone or in hybrid usage, the heterogeneity and hierarchy that will be present in future systems. The Seminar will present the model, examples and results on the different implementations we provide: SMPSs for homogeneous multicores, CellSs for the Cell chip, GPUSs for GPUs, etc. For publications, documentation and access to the distribution look at bsc.es/smpsuperscalar and bsc.es/cellsuperscalar The seminar will be delivered by Prof. Jesus Labarta, Director of the Computer Sciences department of the Barcelona Supercomputing Center (http:/www.bsc.es/) A tentative agenda is given below although the seminar will be very flexible adapting to the questions and interest of the audience. Schedule: 10:00 ΓÇô 10:30: Motivation and basic concepts 10:30 ΓÇô 11:00: Some examples 11:00 ΓÇô 12:00: Basic Implementations: CellSs, GPUSs, SMPSs, Locality. 12:00 ΓÇô 01:00: Lunch Break 01:00 ΓÇô 02:00: Aliased Regions, Hybrid MPI+SMPSs, Load Balance 02:00 ΓÇô 03:00: More examples and discussion