Tensor contractions are fundamental computational primitives for highly-accurate coupled cluster methods. By leveraging extreme-scale HPC platforms, the computational requirements of large basis set problems could be made feasible. Scaling digital computing in the post-Moore’s law digital computing would require coping with diverse and challenging architectural trends that are driven by system efficiency constraints, including massive parallelism, complex memory hierarchies, lightweight compute core designs, etc. For instance, most lightweight core designs are inefficient in handling complex software stacks or runtimes and thus work best with low overhead runtime systems. On the other hand, computational scientists favor high-level software abstractions, potentially with sophisticated runtimes, to ease the expression of their computational expression and to improve their productivity. They also prefer to interact with a unified programming abstraction across different architectures and run scales. In this talk, we will discuss bridging the widening gap between these conflicting requirements through software frameworks that leverage multiple programming models while providing a unified interface. We will also discuss the role of software stacks in maintaining performance portability across different architectures and computational scales. Our case study is the Libtensor framework, for which we will present performance analyses on leading supercomputing facilities and discuss the requirements for future architectures.

Scaling tensor contractions in the post Moore’s law computing era

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.