When to use Intel CMK vs native CPU manager?

[Prashanth B]

Hey all,

I'm looking for some tips on when to use Intel CMK over the stock Kubernetes CPU manager. I saw the table in section 2.3 of the Intel CMK guide, but it doesn't call out the points below. 

* AFAICT it seems as if the native Kubernetes solution CAN guarantee cpu pinning/isolation for containers, but CAN'T guarantee that all the threads in a container don't get moved around by the kernel scheduler - which intel cmk can, because those cores are isolcpus. 

* I think another limitation of the native CPU manager, according to this writeup, is that while you get the cores you need they might be from 2 different sockets? Is this solvable with the topology manager? 

Does anyone have a better comparison of the two? I'm a little new to the resource scheduling space (not new to Kubernetes), so apologies in advance if my questions are basic from that perspective. 

Thanks,

Prashanth