Original issue: https://charm.cs.illinois.edu/redmine/issues/2

The current system for reductions buffers input messages at each node of the tree until every input has been received. For many kinds of reductions, this is wasteful, because they can be processed continuously (e.g. logical, int, min/max), or they could be processed in some smaller increment, forming a tree over the messages with good accuracy (e.g. FP addition).