Whats nice is that people are thinking about supercomputers again, even if they insist on calling them "the cloud".
First things first, look up beowulf (http://en.wikipedia.org/wiki/Beowulf_cluster) which is a suite of tools that implents a multi machine scheduler, and message passing interface. Whats nice is that if one host is overloaded, it can migrate the process to another. (however I'm not sure what performance is like nowadays)
In the world of VFX for movies, we've been dealing with schedulers for years. Programmes like alfred, tractor, qube and deadline can dispatch tasks, and deal with dependencies at massive scale.
The first thing to note is that "DCOS" is really a discrete set of parts; a scheduler, machine state enforcement, network config, Storage, and the underlying OS.
With careful planning, the state enforcement tool (puppet and the like) can take care of all of these tasks except a global scheduler.
The beuty of the VFX scheduler is that they understand dependencies really well. (I need x to complete before I run y, I need feature z to run) A lot of newer schedulers really don't understand this concept well.