SCCS Colloquium - Apr 24, 2019

Date:	April 24, 2019
Room:	02.07.023
Time:	15:00 - 16:00

Stephan Pirner: Profiling of a Distributed Task Stealing Implementation in the Parallel Adaptive Mesh Refinement Framework sam(oa)2

This is a Bachelor's thesis submission talk. Stephan is advised by Philipp Samfass.

While easier access to Multi-core architecture and distributed memory drives the development of parallel applications further, load balancing mechanisms for this technology have to be developed. For this purpose, the adaptive mesh refinement framework sam(oa)2 has been developed. While the common approach supposes that an equal distribution of the work on the ranks results in the best runtime, imbalances between nodes are not minded. To get rid of these imbalances a task stealing feature has been implemented in sam(oa)2, this feature provides the possibility to reactively steal a task from another rank on demand in distributed memory. This thesis aims to profile the efficiency of the strategy standing alone and in combination with the a state-of-the-art predictive chains-on-chains partitioning load balancing mechanism. Besides, the chains on chains partitioning is compared in its efficiency to distribute the load among ranks. For this purpose, an integrated profiler is implemented in sam(oa)2 and several experiments are conducted during this work, which will provide deeper information about the load balancing strategies of sam(oa)2.

Keywords: samoa, profiler

Language: English