Allowing Edge Weights for the initial workload partitioning #588
Conversation
…w partitioning in respect to communicaiton cost. The user now has to set NodeWeightModelTypeId and EdgeWeightModelTypeId instead of the LtsWeights arguments that was aviable before, which is removed
…nd edge weight strategies
|
I have updated everything according to the comments |
|
I have tried to apply every comment you have mentioned, I will add the comment for the functions now and make another commit. Other than that I have used the format file, so it should be nice to look at. Regarding the parameters I need to change the documentation a little because in psuedocode we combine with Approximate Communication but actualyl due to the implementation of the constraints we need different EdgeWeightModelTypeIds that implement Communicaiton Approximation + plus a little bit extra, that should be in the documentation right now |
|
I've tested on several setup, and it run through, (at least) without performance losses |
|
I am glad to hear that, right now I have some other urgent tasks to attend to, I will continue with the proposed fixes for the PR as soon as possible |
Hi, it is the essentially the same pull request but I have added the documentation for the edge and vertex weights. In a foolish mistake I have closed the last pull request, so I am opening a new one.
The text from before:
"
Adding edge weights improve the performance in most cases.
To allow this the old LtsWeights class is changed to be a class that contains NodeWeightModel and EdgeWeightModel classes as members, which respectively compute the vertex and edge cost models, depending on the given input for the strategy.
The old LtsWeights parameter that would be defined in .par file is not allowed anymore, user has to define NodeWeightModelTypeId and EdgeWeightModelTypeId.
This change also needs my PR request on PUML2, as we need to get the dual graph representation of the mesh to be able to compute edges. And we propagate the edge weights to metis, and the function to partition the graph will need the accept edgeweights as an additional parameter.
"
Now with documentation to show what it changes! Parameters are updated as well.
If this PR is merged, we also need to link to Ravil's multi constraint branch, my PR is still open there. The link:
TUM-I5/PUML2#5