I'm not convinced that Hurd's translator idea is a hack, or that it's doing it in the wrong layer.
That every program that presents FTP on my system (GNOME, KDE, lftp, Emacs/TRAMP) has to implement its own VFS is a hack, so is having to use iptables or another forwarder if I want to run a non-root daemon on port <1024. And that I have to recompile and reboot if I want to enable some minor feature in the kernel.
EDIT: The last two paragraphs are redundant because I misread rbanffy's reply, and I oddly can't reply to his new one so I'm putting this here.
This is essentially treating the hypervisor as an exokernel operating system. And it seems to become more common.
Perhaps we are really coming to exokernels in a very roundabout way.
I see exokernels as the future, but I'm slightly biased, as I'm involved with one...
Still an active project as far as I know. Mick Jordan has as of late been hacking on Maxine and I've seen GuestVM related commmits in Maxine.
> However, it makes no sense when you start development to try to guess what calls you might end up using.
I agree. That's why you will be using libraries to duplicate what your OS did for you. The Exokernel guys call this a library OS. Libraries are already quite good at providing abstractions.
The difference to a conventional architecture is, that the kernel only provides secure multiplexing between application on the same machine. But the abstractions from the hardware are provided by some libraries. It's easier to experiment with libraries than with operating systems.
The MIT-Exokernel-Operating-System page (http://pdos.csail.mit.edu/exo.html) explains this in more detail. If you look at the first picture, and replace `Exokernel' with `hypervisor', you get modern virtualisation.
I am uncomfortable having what amount to be sci-fi versions of Unix and VMS to chose from. What would be a cutting-edge, legacy-free OS these days?
Though sadly this page is from 1998. But there are some people who are still working on Exokernels. You can meet them here at HN.
Hurd is late because there isn't any need for it anymore, we now have plenty of free kernels. And GNU isn't sponsoring it because of that, just like it isn't sponsoring a GNU MTA.
But work continues because a few people are still interested. There's a lot of neat ideas in Hurd. You get a lot of nice things once you move most of the kernel to user-space.
The best example of this is custom user-mounted filesystems (translators). You can e.g. mount a remote FTP site on Hurd in ~/ftp without being a superuser, this is something that GNOME and KDE implement on their own because users can't do it on other systems without special permissions.
Linux has been accumulating a lot of microkernel-like features over the years. E.g. dynamic loading of device drivers, FUSE etc. In time it'll probably usurp every useful Hurd feature worth having.