First International Workshop on Performance Portable Programming Models for Accelerators (P^3MA)
Frankfurt, Germany
June 23, 2016
Program
Keynote Speaker
Simon D. Hammond
Scalable Computer Architecture Group
Sandia National Laboratories
Balancing Productivity, Portability and Performance - The Challenge for Programming Models at Exascale
ABSTRACT: Architectures for high performance scientific computers are growing increasingly complex. When selecting hardware for a supercomputer the choices may include general-purpose GPUs, conventional multi-core processors, hybrid processors which blend a GPU and processor cores into a single device and most recently high-threaded many-core chips. While the range of hardware choices is good for supercomputer designers, it poses a significantly challenge for application developers who need to write, optimize and support applications that can be migrated between many machines. This is even more challenging when users demand high performance portability.
In this talk I will present results from early evaluation studies at Sandia using the Kokkos programming model compared to solutions such as directives. Our implementations successfully run on diverse architectures including Intel Haswell and IBM POWER8 processors, NVIDIA GPUs, Knights Corner accelerators and processors from ARM licensees. Performance varies significantly reflecting how well the runtimes are tuned by vendors or our development teams. Results will also be provided which evaluate the programmer productivity associated with each approach, reflecting some of the underlying long term software engineering cost with selecting one model over another.
Finally, I will present some thoughts and experiences from five years of porting code to novel test-bed architectures and what lies ahead in the future as we embark on planning for the Department of Energy's Exascale Computing Project due for deployment in 2023.
BIOGRAPHY: Simon D. Hammond is a research scientist in the Scalable Computer Architecture group at Sandia National Laboratories. His primary research focuses on enabling codesign activities for the laboratories using tools such as the Mantevo mini-application suite, the Structural Simulation Toolkit (SST), the NNSA/ASC Advanced Architecture test beds and several other projects of which he is a member. This work includes activities with the Office of Science Materials in Extreme Environments (ExMatEx) and Combustion (ExaCT) codesign centers as well as the Sandia NNSA/ASC Codesign project focused on large scale production-quality engineering applications and mathematical libraries. He has research interests in high performance compilers (particularly vectorization strategies), dynamic runtimes with a focus on the just-in-time compilation and dynamic execution of code, the C++ language standard, efficient message passing, high performance numerical mathematics libraries and the analysis, modeling and optimization of large-scale parallel scientific or analytics applications.