You're missing the point. Supercomputer networks tend to be things like infiniband, where you're (usually) not using TCP and the whole point is to enable thing

You're missing the point. Supercomputer networks tend to be things like infiniband, where you're (usually) not using TCP and the whole point is to enable things like RDMA (yes, that stands for remote direct memory access). IPoIB is available but you pay a performance penalty; on Cray's interconnect the TCP emulation performance has traditionally been really bad. Google's approach (to anything) is irrelevant to supercomputing, where users' program execution is managed by schedulers and the supercomputer is treated as one unit, despite being made up of many potentially-autonomous nodes. Just about all of the tech you've namedropped is completely unrelated to the manner in which HPC code networks. Having said that, nobody uses rsh for performance, because if you really want performance you have to skip TCP altogether. People in HPC who are using rsh are using it because of inertia: the same reason the Top500 is the world's largest repository of csh users.