Skip to content

unibas-dmi-hpc/MapLib

Repository files navigation

MapLib

MapLib provides algorithms for generating mapping of processing elements to processor unities to three 3-D topologies: mesh, torus, and HAEC Box. MapLib implements: (i) Communication- and topology-oblivious mapping strategies, which do not take into account the communication matrices of the applications nor the target processor topologies. These algorithms follow a predetermined node ordering to map all processes to the available nodes and produce deterministic mappings. (ii) Communication- and topology-aware mapping strategies, which consider both communication matrices of the applications and target processor topologies. These algorithms produce different mappings for a given application–system pair.

Paper references

MapLib contains the following process mapping techniques:

Communication- and Topology-Oblivious Mapping:

Peano
Hilbert
Gray
sweep

scan

Communication- and Topology-Aware Mapping:

Bokhari
topo-aware
greedy
FHgreedy
greedyALLC
bipartition

PaCMap

The folder mapping-matters-commMatrices contains all input communication matrices
and
The folder mapping-matters-Mappings contains all mappings that were used in the paper:

Mapping Matters: Application Process Mapping on 3-D Processor Topologies available at: https://arxiv.org/abs/2005.10413

The sub-folder MapLib contains all Python files where mapping.py and mapping_compl.py consist of the implemented mapping algorithms.

To install the program:

  • Run make
  • Run pip3 install --editable .

As an example of how to use the library, one can run the provided wrapper.sh script

Acknowledgments

Acknowledgment to Daniel Besmer and Viacheslav Sharunov for their earlier contribution to the library.