| Author | William R. Tobin |
|---|---|
| Title | The Adaptive Multiscale Simulation Infrastructure |
| Year | 2018 |
| Journal | Ph.D. dissertation |
| Pages | 130 |
| School | Computer Science |
| Institution | Rensselaer Polytechnic Institute |
| Abstract | Increases in the computational power of massively parallel machines have facilitated the implementation and execution of numerical simulations of unprecedented scale and fidelity. When the physical mechanism governing the action of interest in a simulation occurs at a scale orders removed from the desired resolution of the solution, even modern leadership class machines are insufficient to provide full granularity. Multi-scale numerical simulations allow the coupling of multiple scales of interest in order for the multi-scale simulation to reflect important properties of each scale. The whole-clothe implementation of multi-scale simulations targeting massively parallel machines requires a substantial investment of man-hours. Leveraging existing, wellestablished single-scale simulations already in wide use in order to construct new multi-scale systems will allow a reduction in development time and costs to produce novel new multiscale systems. The design and development of the Adaptive Multi-scale Simulation Infrastructure (AMSI) along with the design and implementation of a multi-scale soft biological tissues code is discussed. The parallel characteristics and implications of the AMSI componentcode coupling approach are considered in the context of the soft tissue multi-scale code. Modern simulations on leadership-class machines also leverage dynamic and adaptive capabilities to load balance simulations during execution and reduce time-to-solution. Integration of such codes into a multi-scale system must allow for the use of dynamic load balancing operations while maintaining all inter-scale linkages. The deployment of codes designed for massively parallel machines relies on the leveraging of various component libraries, often fulfilling similar purposes but targeted at specific machine architectures to provide optimal performance. Accessing these component libraries through mechanisms providing for the use of a generic interface over a component code targeted to a particular platform allows for the development of algorithms once and their deployment and optimization on various platforms without the need to explicitly modify code. Implementation of such a mechanism should necessarily introduce minimal overhead into the resulting simulation, with a goal of zero-overhead abstraction of the various component backends. |
| PDF File |  Download |