Sounds a bit like a Beowulf Linux cluster from about 20y ago. What's the major difference between the approach of those.
I'm not convinced a shared memory abstraction is a good thing for distribution across machines. Some Beowulf clusters actually had much lower latency interconnects than ethernet, e.g. Quadrics, but we still needed to carefully optimise the traffic. With a shared memory abstraction, such optimisation is out of our hands.