The Hurd: GNU's quest for the perfect kernel

Are you asking in theory or in practice? Latter depends a lot on the implementation and whether it's a HW driver or just a pure SW driver. :-)

The Hurd distro page had at least this comment:
"You can run several instances of the Hurd in parallel, and debug even critical servers in one Hurd instance with gdb running on another Hurd instance."

(Debugging doesn't succeed if you're single stepping through critical system service, therefore debugger needs to run elsewhere.)

> Does it also demonstrate this IPC overhead micro-kernels are often blamed for?

There was this comment from user (in 2009):
"while X does work, it works very poorly -- it's not only slow and jerky all the time, but also tends to lock up completely."

But I assume the slowness is partly because of missing shared memory:
"Although the POSIX interface is provided, some additional interfaces like POSIX shared memory or semaphores are still under development."

And of course it being a development kernel; not a finished, mature and well optimized one.

Based on the referenced Hurd article and e.g. these Wikipedia articles on Mach + L4 and its derivatives which take into account things not taken into account in Mach design (perforamance, resource handling, security):
http://en.wikipedia.org/wiki/Mach_(kernel)#Second-generation_microkernels
http://en.wikipedia.org/wiki/L4_microkernel_family#History

I would say that Mach derived stuff seems better to move to newer microkernel generation to be able to be relevant to real users (or researchers).

OSX is also partially Mach based, but it's a hybrid, not a pure microkernel:
http://en.wikipedia.org/wiki/XNU