Note that, for Linux, "IPC" is kind of loaded since it tends to refer to TCP/Unix sockets/signals/etc., which aren't used as much as IPC on a microkernel system. The right comparison would probably be with system calls overhead.
I'm misunderstanding something. I interpret "Latency is where µkernels have historically fallen down." as implying that µkernels have historically been worse than Linux. But, that table shows them historically being sub-microsecond.
As I recall, the problem is that raw IPC costs are a red herring. It's possible to get the IPC costs almost arbitrarily small, if you're not actually toting any data or if you don't have memory protection domains separating the components (as in Scout).
If you are toting data around, such as reading or writing to a filesystem server, you have three options:
* Copy it. That's kind of expensive.
* Share it. Copy-on-write magic, for example. Unfortunately, that requires fiddling with the VM system, to set up a mapping between user-space and filesystem-server-space, for example. Fiddling with the VM system can be surprisingly expensive, too.
* Pre-establish a shared-memory buffer. This is what L4 does(?), if I'm reading section 3.2.2 correctly. It may be much better than the other options, I have no experience there.
(Excellent paper, by the way. I'm hoping to get to do something with L4 at some point; microkernels are neat and it seems like it doesn't suck.)
