Collaborative Research: SI2-SSE: A Petascale Numerical Library for Multiscale Phenomena Simulations
Grant
Overview
Affiliation
Other
View All
Overview
abstract
Multiscale phenomena are a grand challenge for theory, simulations and experiments. It is difficult to understand phenomena at both highest and lowest scales, as well as all those in between. This challenge shows up in diverse fields. One of the long-standing problems in cosmology is understanding cosmological structure formation. In computer simulations of this process the challenge is increasing grid resolution while retaining the essential physics. In all-atom molecular dynamics simulations of enzymes the challenge is simulating systems with a large number of atoms while resolving long-range interactions and having sufficiently high throughput. Such simulations are critical in understanding important biological processes and eventually designing new drugs. Another prime example of multiscale phenomena is turbulent flows, a rich and complex subject of great relevance to many of the main technological issues of the day, including climate, energy, and the management of oil and biohazards. Understanding turbulent flows is critical for design of new transportation vehicles, improving efficiency of combustion processes and managing their environment pollution. Here simulations have been historically limited, and remain so, due to extremely high computational cost, even using the high-end computational systems available to researchers today. Reducing this cost, and efficiently using the computational resources, often requires specialized expertise, as well as significant development time and cost many research groups cannot afford. This project will develop a powerful suite of critical software components to provide tools for performing simulations of multiscale phenomena. The suite will implement state-of-the-art techniques for reducing communication cost, which has become the most important contributing factor to the total simulation cost, especially at larger scales. It will provide a flexible set of features that will make it usable in a great number of codes across the disciplines. In particular, the library will include user-friendly interfaces for Fourier transforms, spectral and compact differentiation in three dimensions, in addition to widely used communication routines (transposes, halo exchanges). This combination of emphasis on scalable performance and richness of features makes this suite unique among other libraries in existence today. Given the extraordinary challenge of simulations of multiscale phenomena, this library will provide a realistic path towards the Exascale. The suite will be available under an open source license. User outreach will be undertaken through the project website, mailing list, user surveys and presentations.
