Scale testing

[drmorr0]

We'd like to know how "big" we can actually go with our simulation. I've done ~1000 nodes on my local laptop, but sortof by accident, I haven't done any serious benchmarking or profiling. We'd also like to validate how much space traces actually take up. It would additionally be cool to do some kind of "before/after" scale testing with both the VirtualKubelet implementation and KWOK (see #70).

[Huang-Wei]

According to my conversation with @wzshiming, if the VK impl. sticks to Informer, it has to keep all API objects in memory, and that'd consume a large amount of memory.

In the live demo in KubeCon, at the end of 5k nodes / 10 pods scheduling process, the memory usage of kwok is ~200Mi. (observed from macOS's activity monitor)

Would love to see the comparison.