Critique of Microkernel Architectures – Is Linus Right? (2004) [pdf]

Gosh, I never realized, as a former Keeper of Alphas for scientific work, previously dependent on the GEC 4000 series <https://en.wikipedia.org/wiki/GEC_4000> for physics. I wonder how it compared with OS4000 Nucleus.

The 4000s made competing VAXen look a bit silly for our sort of data acquisition and analysis. Since they were pretty important for my publication record, it always amuses me to see microkernel-ish systems rejected out of hand. A colleague disproved the conventional wisdom that you needed a "spectrum database" to manage the data by just trusting the filesystem to manage directories. More recently I recall userspace filesystems being called generally toys (?), notwhithstanding huge, high performance PVFS2 installations, for instance.

> The DEC Alpha was a powerhouse in its day

Yes, and tragically under-appreciated in my sphere. I tried in vain to get the maintainer of the principal data reduction program in a different field to replace the bottleneck of the disk-based sort (written for a PDP-11). Each image would normally fit in cache and the whole dataset largely in memory, but that somehow wasn't relevant, contributing to "the computer is slow". (I think the sizes were 10MB and 1GB.)

Qubes OS is the hypervisor department's version of the "compartmentalize OS for security" idea. The microkernel dept should come with their own credible version if they want to win.

I wonder how feasible a modern, graphics-capable POSIX OS built on a core of sel4 is?

For the latest video game, with a microkernel (and an IOMMU keeping the video card from trashing memory), when switching to fullscreen mode, your OpenGL library could request that your video game process become the video driver, with no loss of security or stability. That gives you lower overhead access to the video card than a monolithic kernel.

For compute-bound tasks such as bitcoin mining, a small realtime microkernel will give you better instruction cache utilization. (Back in university system architecture class, I ran cache benchmarks under Linux, QNX, and Win2k on my triple-boot desktop, QNX gave me the best cache performance.)

If you need to run a high-performance webserver, if you need absolute performance, run it in-kernel. If not, a microkernel (and a properly configured IOMMU) allows you to safely make your process the network driver as well as filesystem and disk drivers for your dedicated www data drive. Now, you'll want a second network port with its own chipset (perhaps on its own card) for being able to ssh in if the webserver crashes, but you'll have higher performance than a monolithic kernel.

In short, microkernels along with appropriate hardware protections allow you to safely make your device drivers just ordinary libraries that run inside your critical processes, giving you higher performance without sacrificing security or stability.

EDIT: I should also point out that Oracle, Postgress, and other high-performance RDMSes go through a lot of work to do "kernel bypass". A microkernel with minimum cache footprint would better get out of their way and more safely turn over more functionality to the application.

Neither iOS/OS X or NT manage to get the robustness, security or compartmentalization advantages of microkernels. I guess they get modularity and dynamic loading, but so does Linux. So Linux, X and NT are all mongrels, with iOS/OS X having microkernel roots and Linux having monolithic roots, and NT being a ground-up designed mongrel.