Even if the above method worked, loading each grid onto 1 processor/node, it would only alleviate the memory problem so much, because potentially a LOT of the particles can be on a single grid, which will still overload the memory sometimes, so this isn't as good as the parallel HOP's KD tree way of cutting up the particles for load balancing.
Are you saying that your first example of how to read the particles (adapted from what I gave you) ran out of memory? Let me know!
Regarding running out of memory, as particle IO works in yt currently, a grid with lots of particles can still be a problem, even in parallel HOP. If you are in fact running out of memory due to a really heavy grid, we should think about how to address that.
-- Stephen Skory email@example.com http://stephenskory.com/ 510.621.3687 (google voice)