Component-based Parallel PDE Simulation

Modern computational science often involves simulataneously computing multiple large-scale physically coupled simulations. These simulations are complex, and require efficient execution on hundreds to thousands of processors. Furthermore, requirements of simulation may be dramatically different than a companion simulation running in the same computational domain.

The Uintah Computational Framework (UCF) is a set of software components and classes that are designed for parallel simulations on structured AMR grids, providing capabilities such as semi-automatic parallelism, automatic checkpoint/restart, load-balancing mechanisms, resource management, and scheduling. The UCF exposes flexibility in dynamic application structure by adopting an execution model based on software or "macro" dataflow. Computations are expressed as directed acyclic graphs of tasks, each of which consumes some input and produces some output (input of some future task). These inputs and outputs are specified for each patch in a structured grid. Tasks are organized in a UCF data structure called the task graph. The UCF storage abstraction is sufficiently high-level that it can be efficiently mapped onto both message-passing and shared-memory communication mechanisms.

We will discuss the architecture of the system and the algorithms that are used for scheduling and communication. We will discuss the performance of the system on simulations utilizing up to 2000 processors.

This software is the foundation of the University of Utah's Center for Simulation of Accidental Fires and Explosions (C-SAFE), a DOE ASCI Level 1 ASAP Center.