Do you have some design recommendation regarding workitem count for the one node and for the whole network?
Something like comfortable workflow for average computer and if user goes beyond some count then probably need redesign a network in some other way.
I undesrand that each case is induvidual but still ...

I think this is where the schedulers come in play, I have so flows where processing is light and easy and I have it set to the standard 25% of the available CPU cores. But I have flows where some part of it requires a lot of computational power from all the cores because of a multithreaded process in it. Then I add a new schedular for that particular node and limit the slots to 6 or 4.
The way I come up with the right number is trial and error, I have a 64 core machine with 128Gb of ram so I try a few settings to find the sweet spot in terms of machine load and memory consumption. If I do 6 slots and the memory fills up right away, it means the machine will start swapping memory with the disk and it will be slow. So I set it to 4 slots, check the memory doesn't exceed 110GB of total usage and go with that.

So yeah as you said it's highly individual based on machine characteristics.

My question is more about TOP capability. For example, when I have more than 20 000 work items they are slowing down overall navigation noticeable, especially if I have visualized dependencies between them.