Distributed GIS Support

[etatara]

Support GIS / geographic projections in a distributed (multi-process) space in a way that is similar to the existing spatial projections, that supports automatic moving of agents across ranks, ghosting, etc. API should attempt to closely align with the single-process Geography API (#41).

• How to partition. Every partition would have the knowledge about the global partitioning AND features that were split up based on the partitioning.
• There are at least two types of split features:
  • Auto ghost, where we want to understand agent presence in them, so if an agent enters in one partition, they are automatically ghosts in the other partitions.
  • Non-auto ghost features
• Queries originating on partitions will be of 2 types:
  • Within the buffer of the partition
  • Outside of the buffer, where the global map would be used to understand to which partitions the spatial queries should be sent

Beyond partitioning, importance of inputting/outputting shapefiles, and CRS generalization beyond what we do for RS.

Also NC - search for multi-process GIS ABM prior work. Vague memory of reviewing some.

How to partition global space if some features extend over boundaries?

• For static features (e.g., roads) could create new geometry by intersecting with area controlled by specific rank.
• For dynamic features, buffer big enough to hold full feature size. What if size is very large?

Implementation Ideas:

• GEOS (https://libgeos.org/) seems to be the best choice on the C++ side. The Shapely python library uses GEOS, so the underlying geometry API is similar. And GEOS also supports geography operations and spatial indexing, so if we store the agent/geometries entirely on the C++ side, we may end up with an architecture that is close to the existing repast4py space projections.
• We'll probably want to use python for data loading and projection transforms, so the single process API (Single process GIS support #41) can support this through fiona and pyproj.