Task-based programming models: tradeoffs between granularity and computing platforms

Rosa M. Badia, BSC-CNS

Task-based programming defends that provides a simple programming approach that leverages the computing infrastructure and reduces global synchronizations. Since its inception has been extensively applied to different computing platforms. In particular, at BSC we have developed instances of programming models for so diverse infrastructures as the grid, multicore or the Cell processor, clusters and large supercomputers, to the digital continuum or hybrid classic-quantum approaches. However, is the programming model so simple? What the end-users say about it? What are the issues that each computing platform pose? Is an autonomous runtime taking all decisions appreciated by end-users? How can we debug or improve the performance of distributed applications? The talk will present a perspective overview of these topics and our views for the next future.

Fortran and the OpenMP API in a Modern GPU World

Michael Klemm, AMD

Modern day supercomputers are massively parallel, heterogeneous systems that employ accelerators (mostly GPU) to provide additional compute and memory performance to applications. While C/C++, but also Python, gain traction in the HPC domain, Fortran continues to have a large developer base with new high-performance code written every day. In this world, the OpenMP Application Programming Interface is one of the key components to support application developers and their need to write portable and performant code for such systems, especially in the context of large Fortran codes. We will review the evolution of the OpenMP API from its early days in 1997 to the present day and how it supports large scale, heterogeneous applications. We will recap how OpenMP initially supported portable multi-threading (for Fortran) and how it was extended to support task parallelism, single-instruction multiple-data, and heterogeneous computing. We will also shortly touch on future plans for the OpenMP API versions 6.1 and 7.0. The review of OpenMP features will be embedded in the journey of AMD to build the first and second exascale system, based on AMD EPYC(tm) Processors and AMD Instinct(tm) accelerators as well as a modern Fortran compiler based on LLVM Flang. Buckle up and enjoy the ride!

Title tba

Nick Brown, EPCC

Abstract tba