16.01.2013 Views

Lecture Notes in Computational Science and Engineering - Bioserver

Lecture Notes in Computational Science and Engineering - Bioserver

Lecture Notes in Computational Science and Engineering - Bioserver

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

12 Gudula Rünger<br />

Data <strong>and</strong> communication management for adaptive f<strong>in</strong>ite element methods<br />

The efficient parallel implementation of a hexahedral adaptive f<strong>in</strong>ite element<br />

method has been presented <strong>in</strong> [16]. In this approach, the communication is<br />

encapsulated <strong>in</strong> order to provide a general mechanism for repartition<strong>in</strong>g <strong>in</strong><br />

graph based algorithms.<br />

The ma<strong>in</strong> characteristics <strong>in</strong>duc<strong>in</strong>g irregularity to adaptive hexahedral FEM<br />

are adaptively ref<strong>in</strong>ed data structures <strong>and</strong> hang<strong>in</strong>g nodes. A hang<strong>in</strong>g node is<br />

a vertex of an element that can be a mid node of a face or an edge. Such<br />

nodes can occur when hexahedral elements are ref<strong>in</strong>ed irregularly, i.e. when<br />

neighbor<strong>in</strong>g volumes have different levels of ref<strong>in</strong>ement. For a correct numerical<br />

computation, hang<strong>in</strong>g nodes require projections onto nodes of other ref<strong>in</strong>ement<br />

levels dur<strong>in</strong>g the solution process. The adaptive ref<strong>in</strong>ement process<br />

with computations on different ref<strong>in</strong>ement levels creates hierarchies of data<br />

structures <strong>and</strong> requires the explicit storage of these structures <strong>in</strong>clud<strong>in</strong>g their<br />

relations. These characteristics lead to irregular communication behavior <strong>and</strong><br />

load imbalances <strong>in</strong> a parallel program.<br />

The task pool team approach from Sect. 2 provides a concept which allows<br />

any k<strong>in</strong>d of data references or communication patterns, <strong>in</strong>clud<strong>in</strong>g strong irregular<br />

communication. For adaptive FEM, however, the locality of references<br />

<strong>and</strong> communication is slightly different than described for the algorithms presented<br />

<strong>in</strong> the last section.<br />

• The adaptive FEM has a data oriented program structure with dynamically<br />

grow<strong>in</strong>g (or shr<strong>in</strong>k<strong>in</strong>g) data structures. Based on the <strong>in</strong>put data a<br />

suitable <strong>in</strong>itial distribution of data can be chosen at program start.<br />

• The computational effort varies dynamically accord<strong>in</strong>g to the dynamic<br />

behavior of the data. The process is guided by the FEM algorithm with<br />

appropriate error estimation methods.<br />

• The data dependencies are of local character, i.e. there are dependencies to<br />

neighbor<strong>in</strong>g cells <strong>in</strong> the mesh or graph data structures. This leads to local<br />

communication patterns <strong>in</strong> a parallel program with only a few neighbor<strong>in</strong>g<br />

processors.<br />

• For efficient parallel execution, load redistribution is required dur<strong>in</strong>g runtime.<br />

But load <strong>in</strong>creases result<strong>in</strong>g from adaptive ref<strong>in</strong>ement are usually<br />

concentrated on a few processors. Appropriate load redistribution may affect<br />

all processors but are of local nature between neighbor<strong>in</strong>g processors.<br />

• Communication occurs for remote data access when data are needed for<br />

calculation. Also shared data structures have to be ma<strong>in</strong>ta<strong>in</strong>ed consistently<br />

dur<strong>in</strong>g program run. In most cases the exact time of communication is<br />

not known <strong>in</strong> advance. Message sizes or communication partners vary <strong>and</strong><br />

depend on the <strong>in</strong>put. However, the communication partners work synchronously.<br />

A parallel FEM implementation can exploit the locality of references <strong>and</strong><br />

local communication pattern. A suitable data management <strong>and</strong> communication<br />

layer for adaptive three-dimensional hexahedral FEM is presented <strong>in</strong> [18].

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!