SCCS Colloquium - Apr 15, 2020
|Date:||April 15, 2020|
|Room:||Online (password: SCCS)|
|Time:||15:00 - 16:00|
Maximilian Karpfinger: Exploring Load Balancing for Modern Numerical Software
Bachelor's thesis submission talk. Maximilian is advised by Philipp Samfass.
Solving hyperbolic differential equations is a very important field in high performance computing, yet it is not only highly time-demanding, but also critical for resources as memory and energy. The more important is an efficient numerical solution to such problems, since an increase in computation time also leads to an increase in energy consumption. However, these numerical schemes might force load imbalances which cause wait-times. Causes for such inequalities have to be detected and fixed. In this presentation the root for imbalances will be presented and their impact on the total run-time will be investigated. The motivation is to reduce these by implementing load balancing schemes that try to balance load distribution at run-time. Such an approach which is based on MPI-wait-times will be examined in the following for different scenarios in the ExaHyPE engine where an improvement in the overall elapsed wall-time of one third was measured.