ahp-graph

ahp_graph - Python Attributed Hierarchical Port Graph

AHP Graphs - https://arxiv.org/pdf/1802.06492.pdf

Requires Python3.9 at a minimum. There is a "noTyping" branch which works with older versions of Python (like 3.6)

Ubuntu apt packages:

apt install graphviz libgraphviz-dev

RHEL/CentOS yum packages:

yum install graphviz graphviz-devel

Python3 packages:

python3 -m pip install pygraphviz orjson

Useful Tips

• Assembly names are automatically prefixed to the expanded devices names when the graph is flattened
• Besides graph construction, the main features that ahp_graph provides are several outputs
  • SSTGraph.build() will immediately turn your graph into SST components and begin a simulation
  • SSTGraph.write_json() will output the graph along with parameters in JSON format
  • DeviceGraph.write_dot() will output the graph in DOT format and optionally draw SVGs with them
• All output generated by ahp_graph goes into a folder called 'output'
  • This is for the SSTGraph.write_json() and DeviceGraph.write_dot() functions
• SST global parameters are set by adding things to the DeviceGraph attr field
• SST component parameters are set by adding things to the Device attr field
• You can still do normal SST Python things (examples below):
  • sst.setProgramOption()
  • sst.setStatisticLoadLevel()
  • sst.setStatisticOutput()
• SSTGraph.build() returns a dictionary of devices with their name as the key
  • You can use this to enable statistics for devices after they are created
• If you plan to build the graph with ahp_graph partitioning and run with MPI, remember to include the --parallel-load=SINGLE flag to sst
  • --parallel-load by default will work with multiple unique files
  • ahp_graph can generate the graph for each rank using a single file, so it needs the value 'SINGLE'
• By default sst grabs all command line parameters
  • You can pass parameters to the python script by putting -- before your script parameters