The performance penalties aren't intrinsic to microkernels, so much as to Mach because of the heavyweight RPC mechanism and its elaborate port right and type checking of messages. L4 calls are actually really fast precisely because such policy is left to userland.

Debian GNU/Hurd runs really well under virtualization, btw.

I certainly don't want to rehash some of the arguments about it, because I wasn't a fan of them at the time...but I was referring to the context switch penalty of running multiple system level services is separate processes.

Mind you, I've always thought that argument was specious for several reasons...but in a naive implementation breaking a kernel unto multiple processes would have to have higher context switch overhead.

I've always though, though, that not everything needs to be process isolated, and using more of a message pass system could probably overcome context switch overhead in multiprocessor system's

EDIT: GOD DAAMN PHONE. I think that the reliability gains may more than make up for any perf problems.

The toll on performance is the price of stability and modularity.

So in a way micro is to RISC ̶C̶I̶S̶C what monolithic is to CISC ̶R̶I̶S̶C.

Umm... if anything I'd say it's the other way around.

Ha, right, I accidentally swapped the terms.