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.

10 Gudula Rünger<br />

The hierarchical radiosity method is one of the most challeng<strong>in</strong>g problems<br />

concern<strong>in</strong>g irregularity of data access <strong>and</strong> dynamic behavior of load <strong>and</strong><br />

communication. Communication is required for remote data access as well as<br />

for load balanc<strong>in</strong>g between nodes. Task pool teams provide an appropriate<br />

tool for h<strong>and</strong>l<strong>in</strong>g load imbalances on SMP nodes as well as on entire cluster<br />

platforms. Load balance between SMP nodes is realized explicitly offer<strong>in</strong>g<br />

additional possibilities for optimiz<strong>in</strong>g the expensive redistribution of data or<br />

migration of tasks.<br />

The programm<strong>in</strong>g model of task pool teams also has been used to realize<br />

different parallel variants of the simulation of diffusion processes us<strong>in</strong>g r<strong>and</strong>om<br />

Sierp<strong>in</strong>ski Carpets [9,20]. In this application the task pool team approach<br />

is especially suitable to efficiently realize the boundary update phase of the<br />

algorithm which is necessary after time steps. The first variant has a synchronous<br />

parallel update phase which exploits the specific communication thread<br />

provided by task pool teams; the exchange of boundary <strong>in</strong>formation is started<br />

by collective communication of the worker threads <strong>and</strong> the communication<br />

thread is responsible for the process<strong>in</strong>g of the data received. The second implementation<br />

variant realizes an asynchronous boundary-update us<strong>in</strong>g only<br />

the task pool team’s communication mechanism. Measurements of the execution<br />

time on a Xeon-Cluster with SCI network <strong>and</strong> a Beowulf cluster with<br />

Fast-Ethernet show that the synchronous approach is slightly better.<br />

The central aim of us<strong>in</strong>g task pool teams is to support communication<br />

protocols that are suitable for dynamic <strong>and</strong> asynchronous communication,<br />

but which do not rely on specific attributes of the MPI library such as thread<br />

safety [17]. Thus, the implementation provides great flexibility concern<strong>in</strong>g the<br />

underly<strong>in</strong>g communication libraries, the parallel platform used, <strong>and</strong> specific<br />

application algorithms.<br />

3 Data <strong>and</strong> communication management<br />

for adaptive algorithms<br />

Adaptive algorithms adjust their behavior accord<strong>in</strong>g to the specified properties<br />

of the result to be computed. This <strong>in</strong>cludes an adaption of computation<br />

<strong>and</strong>/or data <strong>and</strong> is usual guided by a required precision of a solution. A typical<br />

example is the adaptive f<strong>in</strong>ite element method for solv<strong>in</strong>g partial differential<br />

equations numerically.<br />

The f<strong>in</strong>ite element method uses a discretization of the physical space <strong>in</strong>to<br />

a mesh of f<strong>in</strong>ite elements. The numerical solution of the given problem is<br />

then computed iteratively on the mesh. The result<strong>in</strong>g numerical algorithm<br />

has ma<strong>in</strong>ly local dependencies s<strong>in</strong>ce only values stored for neighbor<strong>in</strong>g mesh<br />

cells are used for the computation. Because of these local dependencies, a<br />

useful parallel implementation strategy exploits a decomposition of the mesh<br />

<strong>in</strong>to parts of equal size <strong>in</strong> order to obta<strong>in</strong> load balance [3]. Communication is

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

Saved successfully!

Ooh no, something went wrong!