Partitioning of Structured Adaptive Grid Hierarchies
Funded by :
Mathematics, Information and Computer Science Division
and
|
|
|---|
|
|---|
To enable realistic simulations of critical three-dimensional phenomena, the challenge --- and the goal of this project --- is to enable consistently effective use of all involved resources. Significantly improving the scalability of large structured adaptive mesh refinement (SAMR) applications requires sophisticated capabilities for utilizing the underlying parallel computer's resources in the most efficient way. The meta-partitioner (Figure 1) is a tool providing such capabilities.
The primary motivation for the present project is twofold, viz. (1) No single partitioning technique performs the best for all applications and computer systems, and (2) No established partitioning technique copes efficiently with large-scale SAMR applications with deep grid hierarchies executing on general parallel computers.
We are developing a hybrid partitioner implemented as the software tool Nature+Fable (Figure 2), involving both domain-based and patch-based components. All involved parts are engineered as components conforming to the meta-partitioner. Thus, they are capable of changing behavior to adapt to the dynamics of SAMR applications. Moreover, they are ready to operate in a hierarchical and incremental scheme necessary for supercomputer simulations. Such a scheme localizes communications and avoids costly multi-processor synchronizations and long partitioning times.
Nature+Fable separates homogeneous, un-refined (Hue) and complex, refined (Core) domains of the grid hierarchy and clusters refinement levels into bi-levels. The Hues contain the portions of the grid hierarchy without refinements; consequently they contain only parts of the base grid (refinement level 0). The Cores contain the portions of the grid where refinements are present. The Cores are separated from the Hues in a strictly domain-based fashion, meaning that each Core contains a portion of the base grid and all its overlaid, refined grids. Expert blocking algorithms are used for the Hues. The Cores are subjected to a coarse partitioning, creating ``easy-to-block'' bi-levels. Then the same expert algorithms operating on the Hues are used for these bi-levels.