From what I can figure out, the secret sauce is that instead of CFS enumerating 150-200 processes to decide who gets resources, it instead enumerates just 3-4 cgroups, reducing the scheduling load by a factor of 50.