I have never claimed to be knowledgable of map-reduce. I'm hoping to learn here if I ever get anything interesting done.
If the documents arrive serially, map-reduce is uninteresting to scale across documents. However, there are the multiple hash tables. Now, for with parameters from Petrovic, for a large (1Mdoc) store, you have ~72 tables. Yes, you could put 72 tables on 72 nodes: map sends things to them, reduce collates the results. I've never seen a hadoop 'thing' that has permanent in-memory state like this. I'm not sure where memory mapping comes into the picture. If folks are game, I'll poke the question of contribution some more. Of course, this is as always a question about overall throughput versus respond time. my thinking about this is colored by caring about immediate response. Up to some scale, the money to buy a cluster will buy a whole lot of cores and memory in one machine, and a whole lot of very-low-overhead paralllelism as a result.
