Scott Quinn wrote: I would still suspect a mk-based design to fall short on performance. E.g., I believe Mach-US and -UX were notably slower implementations than the monolithic bsdss running on the same mk. If you look at "true believer" mk architectures, you see things significantly decomposed and *lots* of little servers involved in *everything* (e.g., even resolving a path in a filesystem involves lots of abstraction -- directories are active objects, not passive "name lists", etc.). And, each one of these incurs a significant performance hit because "communication" isn't *just* "pass a parameter list on the stack". OTOH, in embedded designs where things like name resolution and object identification are largely static (and can happen *once* or *infrequently*), that overhead falls to a more manageable level (e.g., the cost of setting up the equivalent of "pipes" is born once).