Banks, multi ports and special connection patterns

[cmnrd]

This issue is closely related to #190. Since this issue grew to be quite extensive, I figured it would be best to break certain aspects down into a separate issue.

So while discussing the issue in #190 we often said that there is no use-case for referencing individual ports and that we should always connect entire banks and/or multi ports. However, as it turns out there is at least one use-case... meet the All-Pairs Shortest Path (APSP) benchmark from the Savina suite. The benchmark essentially consists of a grid of workers, where each worker solves a part of the problem and sends updates to its neighbors. A grid of 4x4 workers looks like this:

where each of the rectangle represents a worker actor/reactor and each worker is connected to its neighbors.

So how can such a grid be build in Lingua Franca elegantly? Of course, all the workers can be placed in a bank, but then there is no way to draw individual connections. Even if there would be such a mechanism, listing all the connections explicitly is not a real solution.

In order to solve this issues properly, I think we would need an expressive syntax in LF. But I don't think that's the way to go, as we will only get more problems (similar to the parameterized widths). Maybe it is a good alternative to use the expressiveness of our target languages instead. We could expose the connection mechanism in a callable function and allow drawing of connection in a startup mutation. Saroush also proposed, that the target code could generate an adjacency matrix, that then is respected while creating reactions.

[edwardalee]

Yes, this could be another higher-order pattern similar to the ScatterGather example. It could be realized today in a manner similar to ScatterGather, but a better approach would be to first abstract what ScatterGather does into an API. Today, ScatterGather reaches directly into the guts of the C implementation.

[lhstrh]

I've been building the API for mutations in the TypeScript runtime. I'll take a stab at implementing All-Pairs Shortest Path using a startup mutation.

[cmnrd]

Yes, a higher level function could help to solve this. It just should be general enough so that various connection patterns and not just a plain grid can be created. I am not sure how this could be achieved. I think you are right, and we should take small steps and find a general solution for ScatterGather first.

[cmnrd]

Closing this. #406 proves that our current connection syntax is sufficient for covering all benchmarks from the Savina suite that do not support mutations.