buffering updates to ygm::container types #117
Description
I am working with a usecase where I am creating a ygm container, e.g. ygm::container::map<std::size_t, std::vector<float>>, from a series of updates to the container. I might have many updates that affect the same key-value pair, which take the form of creating an std::vector<float> and adding it element-wise to the stored key-vector. I tried naïvely implementing this workflow using a separate message for each update, but this actually scales very poorly as the number of keys gets large. I found that buffering the updates locally to an std::unordered_map<std::size_t, std::vector<float>> and only sending one off-rank message per off-rank key resulted in much better scaling.
Would it be worthwhile for us to consider adding an async_buffered_visit method to our key-value store containers, such as ygm::container::map, and ygm::container::array? Such a method could be patterned off of async_visit while accepting an additional lambda function that "combines" two visits to the same key. It might be challenging to do this in a fully general way, however.