Cross Region Unikernels at the Edge
nanovms.com
nanovms.com
It is important to note that different context switches have different costs some of which we have removed. It isn't just one thing. A hundred different processes running on 1 vcpu has a huge perf cost.
It is still possible to get much greater performance this way without having to cut corners. We routinely benchmark applications such as go/rust performing significantly better because of this.
As for needing to change page tables - you need to deal with that regardless.
To dig deeper there is an awful lot going on in say linux to deal with the fact that you have many different programs and (users) running on the same system. Whether it is shared memory or the scheduler itself - there is a ton of work that goes into that and by explicitly not supporting that you can make lots of gains.
I don't think we are trying to promote a purist approach to this - we're trying to be pragmatic.
Call it a duokernel?