|
The Open FUSION Toolkit 1.0.0-8905cc5
Modeling tools for plasma and fusion research and engineering
|
Mesh construction in OFT consist of four phases: 1) loading, where the initial mesh is imported and any associated geometry information is loaded and linked to the internal representation. 2) Local construction, where the local grid is refined and the mesh graph and linkage lists are created. 3) Decomposition, where the grid is partitioned and scattered amongst the available processors. 4) Global linking, where seam information is assembled linking common mesh entities on adjacent domains.
The specific operations performed during import are specific to the mesh type being used. However, the general actions performed during this phase are outlined here. The initial mesh is read from file and loaded into the internal representation. Currently, OFT does not support domain identifiers so any block information is discarded at import. An internal geometry representation is then constructed, which is unique to each CAD type.
In most cases the Native mesh format should be used, along with included python conversion scripts to map from other formats. Some direct legacy interfaces are also available, but should only be used after interacting with an OFT developer.
The imported mesh generally contains only a cell list which defines the tetrahedra in terms of the vertices. Finite element methods requires counts and linkage information for the remaining entities as well. During local construction this information is assembled by traversing the mesh in different ways to identify and link vertices, edges, faces, and cells to one another. During this phase the boundary mesh is also assembled and linked to volume mesh. This process is the same on all levels regardless of decomposition, all global context is contained in seam and global indices which are assembled and updated during global linkage construction.
In order to accommodate distributed memory environments the mesh is partitioned using the METIS library. This library partitions the mesh by attempting to evenly balance the number of cells in each partition while minimizing the number of edges which link partitions. This can be think of as roughly load balancing each domain, as the work should be proportional to the number of cells, while minimizing the communication required, which scales with the number of elements present on more than one partition. The partitioning is done on the master process and communicated to all other processes.
Once the grid has been partitioned seam information linking adjacent processors must be constructed and maintained. This is done by locating boundary points on adjacent domains which have the same global index. A consistent global index is maintained across all processors by refining the global indices created on the base mesh, which is constructed on each process. Adjacent domains are known by analyzing the grid partitioning to locate domains which share at least one vertex.
1.9.8