 |
|
|
|
1999 MCS Divisional Seminars & Colloquia |
|
Optimizing Data Locality in Distributed PDE SimulationsDino Konstantopoulos
|
|
| Abstract | Computing for scientific applications such as finite element modeling is often associated with low-level code, close to machine representation and away from higher-level object-oriented concepts. However, with distributed computing environments that allow a simple laptop to reach out across the network and become a supercomputer, object abstractions are required in order to network different machine architectures and to reuse components instead of reinventing them. As we design these objects with the network in mind, driving touchstones are algorithms for coarse-grained parallelization and data that is well localized (improved loop vectorization and cache access patterns) with minimal latency in machine to machine communication. To highlight such issues, I present the design of a novel Navier-Stokes solver and its associated data container. The algorithm borrows concepts from Lattice-Gas solvers in order to improve stability and leverage state of the art multimedia extension libraries (Sun's VIZ and Intel's MMX). The methods are separated from data containers that are specifically designed for cross-machine distributed simulations. The overall architecture is of a modular type, in order to accommodate the seamless plug-in of new algorithms, and extensible to current industry technologies in distributed computing that can be pressed into service of scientific applications. |
| [MCS | Research | Resources | People | Collaboration | Software | Publications | Information] | |||
|
