The super-carrier + MMscs combo helps with that memory allocation noise. I’d also look at scheduler binding though — on 128 cores the default placement can lead to a lot of thread bouncing between cores, which compounds the context switching problem. Worth looking into if you haven’t already.
Right, definitely have some custom settings where I’m at in this regard 
This turned out to be an instrumentation problem. Just about all cores running at 99%, no high sys load, abnormal context switching, etc, just happily doing a bunch of work during peak load time (re-hydration).