Feature/graph of grid: well envelope

[michal-toth]

This PR adds an interface for using the function addNeighboringCellsToWells of GraphOfGrid that during partitioning merges each well with a layer(s) of cells surrounding it, in order to keep the well further from the subdomain boundary. To use this option, flow must be called with parameter --zoltan-params=*.json taking a json file. The json file then must contain the keyword EnvelopeWellLayers with the number of layers (default 0). The feature is supported only by the partitioner using the GraphOfGrid (--partition-method=3).

In addition, the communication of the well cells (that are not visible to the partitioner) was streamlined, and we added a few more unit tests.

[michal-toth]

jenkins build this please

[blattms]

Looks good. A few comments...

[michal-toth]

jenkins build this please

[blattms]

Thanks a lot. In general this looks like a very good improvement.

Fist batch of comments. I still need to look at the two test files.

Hopefully there are more useful comments than nitpicking on white space issues. 😅

[michal-toth]

jenkins build this please

[blattms]

Thanks for all the changes.
I will insist on the exception, though.
The arguments on performance might be premature optimization. Hence maybe we need to do some realistic benchmarking to persuade us.

[michal-toth]

After checking the performance of the unordered_map in the extendRootExportList, I looked into map/unordered_map in the GraphOfGrid. There are two instances: the first is the graph itself and stores gID->vertexProperties, while the second is EdgeList inside the vertexProperties and stores gID->EdgeWeight.
Before, we used unordered_map for both, but some testing has shown that using a map for EdgeList (which has usually at most 6 entries) is faster. Using map for the graph itself is worse. Both are about 10% difference in performance. However, using map for both resulted in dramatic performance drop - the compute time doubled. I can not explain why, I tried few more runs after waiting to ensure it is not caused by heat-throttling on my notebook, but got the same results.

[michal-toth]

Testing the map and unordered_map times:

graph	edges	Test:	Easy	Medium	Hard 1	Hard 2
un_map	un_map		15162	14196684	169561690	185451570
un_map	map		13910	12138575	163632397	165122661
map	un_map		16852	15992087	191048511	189642649
map	map		13460	28522892	331562668	315050777

graph is the whole structure in the GraphOfGrid (except wells), it maps global ID to the vertex properties (that include the vertex weight and the map of neighbors called the edgeList). edges is the edgeList, the map of neighbors mapping (neighbors') global ID to the edge weight (which is e.g. the transmissibility between the vertex and its neighbor).
Time is in microseconds, each run parsed the grid 100 times. Only the creation of the graph was considered, callback functions from Zoltan did not enter the calculation. The reserve for the unordered_map of the graph was added later, so the numbers for the best case can be even a little better.

[michal-toth]

jenkins build this please