The Super Instruction Architecture (SIA) is an approach to parallel programming which targets scientific applications that require extremely large, possibly sparse, multidimensional arrays. The original instantiation was ACESIII, a software platform for computational chemistry, in particular highly accurate <i>ab initio</i> electronic structure methods. A key feature of the SIA programming model is to statically partition arrays into blocks which maintain their identity as they move around or are replicated in the system, and then define algorithms in terms of primitive operations on the blocks rather than individual floating point numbers. This approach is supported by a domain specific programming language called SIAL and its runtime system. Aces4 is a from-the-ground up reimplementation of the SIA offering extensions to the DSL, new features including the ability to handle block-sparse arrays, and improved software architecture. A variety of standard and novel quantum chemistry methods have been implemented. The SIA/Aces4 platform is under continuing development to improve performance, exploit new architectures including GPUs and Xeon Phis, add new quantum chemistry methods, and extend to other domains. In this talk, we give an overview of SIA/Aces4, with a focus on the design decisions and enhancements that will enable users to exploit emerging computer architectures.

Emerging systems and the super instruction architecture

254th National Meeting (2017)

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