AuthorMark S. Shephard, Cameron W. Smith, Mark W. Beall
JournalNAFEMS Americas Conference 2014
AbstractIncreasingly industry is relying on massively parallel computers to perform high-fidelity simulations. Although substantial progress has been made on the development of highly scalable methods to model behaviors of interest, the currently available tools fall short of providing engineers with the capabilities needed to execute the massively parallel simulation workflows they want to perform on a daily basis. A key aspect of properly addressing the needed capabilities, particularly on the new generations of heterogeneous parallel computers, is moving from a focus on flop rates to consideration of time of execution and energy consumption for a given level of simulation accuracy. Although the majority of the flops in mesh-based analysis is consumed by the forming of matrix contributions and solving the global system, the process of setting-up and controlling the analysis model is substantial. For example the common estimate in the case of mesh-based methods on workstations is that 85% of the cost is involved with the generation and control of the spatial decomposition of the domain that constitutes the input mesh. In the case of massively parallel simulations with meshes of millions, to billions, of elements this percentage can be higher due to the bottlenecks associated with the data transfer. This clearly indicates that the meshing functions must be fully automated and execute on the same massively parallel computers as the meshbased analysis procedure. In addition, as the complexity of the simulations increase it is also becoming clear that the approximations associated in going from the given problem to the one numerically solved should be automatically controlled to ensure simulation reliability. In most all cases, a posteriori methods that iteratively execute analysis steps followed by model adaptation are required to provide the simulation reliability desired. In addition to providing increased simulation result reliability, adaptive methods can yield the desired levels of accuracy with much less total computation. For example, adaptively defined finite element meshes will regularly have two orders ofmagnitude fewer unknowns than non-adaptive meshes for the same level of solution resolution. However, adaptive methods require more complex, and less structured, data sets and operations and the process of adapting the numerical models is constantly changing the distribution of the data and computations leading to load imbalance that will completely destroy parallel scalability without the regular application of dynamic load balancing. This paper discusses a set on ongoing programs developing parallel adaptive industrial workflows using software components interacting through welldefined open interfaces.
PDF File Download