<p dir="ltr">Hi Srinivas, </p>
<p dir="ltr">thank you for your interest in this project! This week, I am abroad and cannot answer promptly and in great detail. This will get better next week.</p>
<p dir="ltr">If I am not mistaken, you should find the benchmark in dune-grid/dune/grid/test/<a href="http://test-loadbalancing.cc">test-loadbalancing.cc</a>. In addition to the Dune core modules, you need to download and install UG and ParMetis, see the "Installing external libraries" on the Dune website.</p>
<p dir="ltr">Once you have compiled the test, you should try to run it on 1, 2, 4, 8,... processes and check when it crashes. Parallel debugging most likely will bring you to UG's parallel infrastructure DDD: <br>
<a href="http://elib.uni-stuttgart.de/opus/volltexte/1999/426">http://elib.uni-stuttgart.de/opus/volltexte/1999/426</a>/<br>
I am unsure and eager to see what comes afterwards...</p>
<p dir="ltr">I would be happy to receive an application for this project. Improving the current situation would enhance the applicability of Dune's UGGrid manager a lot.</p>
<p dir="ltr">Kind regards<br>
Bernd<br>
</p>
<br><br>---- Srinivas Kumar schrieb ----<br><br><div dir="ltr"><div><div><div><div><strong>Hi ,<br><br></strong></div><strong>My name is Srinivas. I'm a Masters student at the Supercomputer Education Research Center at Indian Institute of Science. </strong><strong><strong>Our center hosts India's fastest supercomputer. <br><br></strong>I've a masters degree in Mathematics and I'm currently pursuing a Masters in Computational Science. In mathematics, I have a done a project in theory of PDE (distributions, hyperfunctions,etc.) and learnt about Finite Element Methods (implemented using C++). I also have worked a lot with MPI, OpenMP, CUDA , Hadoop etc. <br><br></strong></div><strong>I'm interested in the Dynamic Load Balancing project that has been offered by your organization. I'm currently implementing the dynamic load balancing for Breadth-First Search method using MPI. <br><br></strong></div><strong>I need to know more about the benchmark code and what exactly has been done earlier so that I can get clarity about what is expected in this project. <br><br></strong></div><strong></strong><div><div><div><div><div><div><strong>I would love to contribute to this project if given an opportunity.<br></strong></div><div><div><div><div><strong><br></strong></div><div><strong>Thanks,<br><br></strong></div><div><strong>Srinivas.K<br></strong></div><div><strong><br><br><br><br><br><br>Mentors</strong>: Bernd Flemisch<br><strong>Type</strong>: Improve grid data structures<br><strong>Needed skills</strong> Intermediate C/C++, MPI programming, reading other people's code<br><strong>Difficulty</strong>: advanced<br><strong><br>Project description</strong><br><br><br>The
UGGrid class is one of the standard Dune implementations of an
unstructured finite element grid. It delegates most tasks to the <a href="http://www.iwr.uni-heidelberg.de/frame/iwrwikiequipment/software/ug" class="">UG</a>
finite element library, which has a flexible and mature grid
implementation. UGGrid runs on parallel computers, where each processor
only holds a part of the grid. The UG library even allows dynamic load
balancing, where parts of the grid can be moved from one processor to
another. Access to this feature from Dune is flaky, however, and needs
code cleanup, testing, and bugfixing. We have a benchmark code that
triggers a certain sequence of grid movements, which currently only runs
on a very low number of processors. It is the goal of this project to
make this benchmark run with as many processors as possible.<br clear="all"><br><br><br>-- <br><div class="gmail_signature">its now or never</div>
</div></div></div></div></div></div></div></div></div></div>