How can we distribute workloads among compute nodes to make best use of the high speed networks in supercomputers?
Partitioning software such as PaToH and Metis has been known for a while. Such algorithms find good partitions in graphs by minimizing the global size of the cut. However, such solutions are not optimal on the complex network topologies of modern supercomputers.
The goals of this thesis is to develop and implement new strategies that make best use of the available network bandwidth and develop and implement algorithms that partition workloads for modern supercomputers.
- C/C++ programming experience
- Understanding of combinatorial algorithms
- Johannes Langguth
- Xing Cai