You've given me an idea. The slow part of microkernels, and the slow part of pre-emptive multitasking, have the same cause: context switching is comparatively slow. You have to invalidate the cache and let it refill, and you have to load new settings into the MMU.

Now I want to look into ways to reduce that slowness. There's a way for the cache: cache by physical rather than virtual addresses. SPARC did that, I think. As for the MMU, I don't know, but I want to find out.