It's possible to spread a UML instance over multiple hosts in the same way that it's possible to spread a single native kernel image across multiple physical machines in an SSI cluster.
A fairly easy, but easy way to do this is to implement a mechanism for faulting pages from one host to another. So, when a thread on one host accesses a page that's on another host, it would fault, and the fault handler would copy the page across.
This would be highly inefficient. A better way would be to implement an RPC mechanism such as the Compaq SSI work, rather than this DSM.
Having done this, you would have a virtual cluster running over multiple hosts which would be fairly similar to a physical cluster running across multiple physical machines.
This virtual cluster would have access to the combined resources of the hosts in the same way that a physical cluster has access to the combined resources of its members.