Diploma and Master Theses (authored and supervised):
"Efficient Process Mapping for Cartesian Topologies";
Supervisor: J. Träff;
Institute of Computer Engineering, Research Division of Parallel Computing,
final examination: 2020-01-20.
In this thesis we introduce different algorithms for mapping physical processes to Cartesian grids. We assume that processes within this grid communicate with certain neighboring processes as defined by a given stencil. We show that this mapping problem is already NP-complete for two dimensional grids and a very simple isomorphic neighborhood. We compare the current state of solutions in the field of High Performance Computing, specifically the Message Passing Interface (MPI). With his algorithm from 2018, W. D. Gropp showed promising performance, which is why we compare our approaches to his work, as well as MPI's standard behaviour.
For qualitatively comparing concrete mappings, we define fitting optimality criteria, based on the core concept of minimizing inter-computation-node communication.
We benchmarked using MPI_Irecv and MPI_Isend and show that our algorithms can find mappings, where Gropp cannot and that we could match or improve upon his resulting mappings in terms of quality and runtime.
The first algorithm, which we present is similar to other graph partitioning approaches, since it utilizes recursive splitting in order to guarantee a logarithmic runtime wrt. the number of vertices in the Cartesian grid. We also guarantee a quality bound for this algorithm, which becomes better with increasing number of dimensions.
In another implementation-variant, we adapted this algorithm for accommodating differently shaped stencils by weighting dimensions before the recursive splitting depending on how much communication happens across them.
The third approach attempts to find hyperrectangular strips within the Cartesian grid, which are then filled similar to MPI's default row-major rank assignment. Although the theoretical bound of this approach is not as good as the first, this assignment strategy yielded the most compact mappings most of the time. Its main shortcoming, however, is its inability to adapt these strips, depending on different stencils.
Created from the Publication Database of the Vienna University of Technology.