You've mentioned Parascale but not GlusterFS which also combines multiple servers into a namespace but with the difference that they have their own transport mechanism which can take advantage of RDMA. This does mean there's a client portion but it runs in user space via the FUSE interface. This allows for parallism on a data object basis (either file or stripe) instead of just static load balancing at mount time and therefore is more adaptive to changing load conditions than Parascale's approach can be. Also it has the concept of translators inspired by Plan 9 which also inspired it's user space only approach. This allows for a rapid development cycle for new features from the user community as well as the service providers as well as an already existing base of translators that can be combined in many different configuration on a pathname matching basis.

Somehow, I don't see Parascale's attempts at parallism to be any different than DNS round robins across multiple NFS servers, after all existing protocols limits the amount of redirection that can be achieved.

Note: Last I heard, Google was also looking at GlusterFS. There's also a German company whose name escapes me with a similar approach but is intrusive to the kernel.