| (Post author and ZeroTier founder here) I thought about titling this blog announcement "Tenenbaum! Thou art avenged!" but decided to stay away from the link bait. That and I hugely admire Linus. I actually think they were both right. For those who don't know, I am referring to the epic smack down that occurred many years ago between Tenenbaum and Torvalds over microkernels: http://www.oreilly.com/openbook/opensources/book/appa.html I think Tenenbaum was right from a theoretical, academic point of view, but Torvalds was right from a pragmatic ship-it-now point of view. He was also right from an early 1990s point of view. Linus was right because monolithic kernels are easier to build and faster for monolithic system use cases. In the 1990s, individual commodity computer systems were really not fast enough for extreme multi-tenancy beyond maybe multi-purpose use within an organization. This is the sort of multi-tenancy Unix was designed for: a box used by, say, a team or a university department. For this use case a monolithic kernel makes sense for performance and simplicity reasons. The advantages of microkernels just don't matter much, since the box generally has one admin and all the users are trusted. Messing with the kernel is not a big deal. But then Moore's law happened. Now just about anything above a Raspberry Pi is fast enough to host lots of stuff. Big servers with dozens of cores can host hundreds, sometimes even thousands of services. While they only take up single racks, these machines are "main frames" in the classical sense and can easily see diverse work loads belonging to multiple clients. It's incredibly wasteful to single-purpose these machines and let them sit idle. So far we've hacked around this with virtualization mostly, but virtualization is heavy and ugly. Think of it as a thick microkernel that runs a bunch of other whole kernels as services. On huge machines the advantages of microkernels really shine. You can extend the system without having 'root' in a global sense. Crashes of individual system extensions don't bring down the whole machine. Everything is multi-threaded to take advantage of lots and lots of cores. It's even possible to run system services in jails where they only have access to the resources they need so that a compromise (e.g. buffer overflow) in a driver does not permit compromise of the entire system. So in the end I think Moore's Law has turned the debate a bit toward Tenenbaum. But Linux has so much momentum I doubt it's going anywhere. Instead I think we will see Linux evolve toward being more microkernel-y in the areas where this is most desirable. Network virtualization is one such area. Another is security zones. Soon we'll have really secure containers, allowing Docker multi-tenancy, etc. But I wouldn't count out alternatives either. There's an opportunity here for something like Hurd if they could ever get that thing really moving. (Get the Hurd moving... huh huh huh...) Edit: Torvalds was also right because Mach, at least at the time of the debate, was pretty slow. Mach is still not the best microkernel design. See this deck: https://www.cse.unsw.edu.au/~cs9242/04/lectures/lect05b.pdf |