Slashdot Mirror
Get To Know Mach, the Kernel of Mac OS X
An anonymous reader writes "Linux is a kernel, not an operating system. So what is Mac OS X's kernel? The Mach microkernel. The debate around Monolithic (Linux) and Micro (Mach) kernels continues, and there is a great chapter online about the Mach system from the very good book 'Operating System Concepts'. Which design is better? I report, you decide." Warning: link is to a PDF.
commence the dead horse beating. i start : mach owns.
Warning: link is to a PDF"
Um, if you want to warn anyone maybe you should warn the sys admin of the server that hosts the PDF file that you just put a link to on the main page of slashdot. I think they'll care a little more about the super slashdot effect (I'm coining that term for when a non-html file is linked to from slashdot - be it pdf, mpg, avi, etc.) than we will about taking the extra time to load.
... benefits and detriments exist for both monolithic and micro flavors. I doubt a conclusion could ever be made about which one is 'better'... because it all depends on context. "How will the system be used?" "What kind of environment will the system be operating under?" "What are the performance goals of the system?" "What types of hardware will the system(s) need to support?"
Each system has benefits... but they almost always rely on the existence of certain assumptions.
So, even though it uses Mach, you can't call it a Microkernel.
Monolithic translates to no modules, correct?
I can't believe people are modding you up for this.
The Linux kernel is monolithic. Linux modules do not run in user-mode. They are loaded into the kernel proper.
mkLinux was an Apple-sponsored effort to run Linux on Mach. The Linux kernel was modified to run in user-mode; it basically became an executable. In fact, you could run multiple instances of the same kernel (or different kernels) simultaneously.
-mkb
That is monolithic as well, but not using the term in the same way. Monolithic essentially means made from a single piece. This CAN refer to modules as well, as the kernel modules aren't built into the kernel binary, but in the case of monolithic vs. microkernel, it doesn't refer to how the kernel is built. Rather it refers to the execution of the operating system kernel. A modular Linux kernel loads as a single executable that then loads modules into it's process space as needed to do things. This is essentially a monolithic kernel. The OS runs as a single process. Microkernel's have the OS split as seperate processes, mostly outside the core microkernel (which has the job of facilitating message passing between all these processes, and lowlevel process management). The Microkernel may or may not do I/O, sometimes seperate processes do. Hope that helps.
stop spreading the myth that Xnu is a microkernel
Well nope. You can insert newly compiled modules into a previously compiled kernel to get new features (that's how the many proprietory video drivers work, for example.) But those are
a) running in kernel space, not user space
b) communicated with by predefined hooks, rather than a generic message passing interfacing.
That's why linux modules, which are superficially like elements of a microkernel, are not really like them at all.
Athletic Scholarships to universities make as much sense as academic scholarships to sports teams.
That's a statement not a criticism by the way ;-)
In general, monolithic kernels run in a single address space and use direct procedure calls / variable accesses to pass data and control flow between subsystems. This is true even if they support loadable modules (like Linux). Any driver or other subsystem in your kernel can (if it wants) access any other part of the kernel.
Although Mach itself is a microkernel, the "xnu" kernel which Darwin / MacOS X uses also hosts other components *in the same address space*. Some of the subsystems (e.g. the BSD subsystem) are large and resemble monolithic systems themselves. The overall system is not a "pure" microkernel, with lots of code moved out of privileged mode. Equally, it's not quite like a traditional monolithic UNIX because of the use of Mach and the other Darwin-specific components (e.g. a (relatively?) stable binary interface for drivers).
as we switch Steve Job's kernel beans with new Roaster's Light Linux Beans.
Now, let's see if he notices!
The kernel that Apple uses in OS X is called XNU.
It uses code FROM Mach, but it is not Mach and it is not a Microkernel. NT (NT 4.0, NT 5.0 (win2k), NT 5.1 (WinXP) does not use a Microkerenl either.
The only OS that I know of that actually uses a Microkernel is GNU/Hurd.
The OS X kernel, called XNU, is a mixture of *BSD kernel code and Mach kernel code.
Yes, yes, there was a ancient debate between Linus and that other guy about Micro- vs Macro-kernels, and guess what, Linus was right.
Apple does NOT use a microkernel. It does not use Mach. It is BASED on Mach. It would be considured a kludge compared to Linux or FreeBSD, but it works out fine.
Similar to how Mustangs are based on Ford Falcons and Granadas from the late 70's and early 80's. Those cars were as much as a failure as Mach, however the Mustang is flashy and many people desire it. So go figure.
Although interesting, Mach was developed at a university and shows a huge number of problems as a result. Notably performance is terrible, due largely to the IPC performance. When people actually tried the "collection of servers" operating system in Mach 3, it was clear it was simply not a workable solution. Workplace OS, Star Trek and any number of other OS's died as a result.
What's sad about this is that the failure of Mach tainted ALL ukernels. By the mid-1990s the idea was basically dead. But what an idea! Don't have your machine on a network? Simply don't run the network program. Using a diskless system? Don't run the disk server. Don't want _VM_... no problem. You can use the exact same OS image to build anything from a minimal OS for a handheld to a full-blown multi-machine cluster, without even compiling. No pluggable kernels, no shared libraries, no stackable file systems, nothing but top and ls.
But it just didn't work. IIRC performance of a Unix app on a truly collection-of-servers Mach was 56% slower than BSD. Unusable. Of course you can compile the entire thing into a single app, the "co-located servers" idea, but then all the advantages of Mach go away, every single one.
Now, given this, the question has to be asked: why anyone would still use it? Don't get me wrong, there are real advantages to Mach, notably for Apple who ship a number of multiprocessor machines. But the same support can be added to monokernals. Likewise Apple's version has support for soft realtime, which has also been added to monokernels. So in the end the Mac runs slower than it could, and I am hard pressed to find an upside.
Of course it didn't have to be this way. The problems in Mach led from the development process, not the concepts within. As L4 shows, it is possible to make a cross-platform IPC system that is not a serious drag on performance. And Sun's Spring went further than anyone, really re-writing the entire OS into something I find really interesting, and still providing fast Unix at the same time. I'd love to see someone build Mac OS X on Spring...
> QNX uses shared memory to pass messages
So does Mach, and it's slow. I've never seen real-world measures to suggest that QnX is fast. All we know is that the performance of the OS itself is good, and that's a VERY DIFFERENT measure.
The slow performance is due to a number of problems:
1) not all MMU's are really suited to this task. Many are slower to set up than just copying the memory around. Sun found this to be at around 5k, below that, it was faster to just copy memory physically.
2) MMUs/VM are based on pages, 2 or 4k typically. Thus passing in a single 32-bit int parameter causes big page hits. You can tune this out, but it's still annoying.
3) Each copy takes TWO context switches - one to switch into the kernel to copy the memory across ports, another back out to the called program. This means that even the simplest "system calls" are twice as slow as under a monokernel, AT BEST.
4) Additionally the data has to be examined to see if it contains ports being passed around, and if so, they have to be translated because the port codes are private to a program (and thus different in the other one).
5) Using mapped memory ignores all the hardware specific solutions to these problems, many of which are built into modern processors.
It's exactly the sort of one-size-fits-all solution that you'd expect from a research project. One that doesn't work in the real world. One that should have been replaced, and was in L4, Spring, etc.
For instance, Spring included three different IPC systems, each tuned to certain types of data, each used in different ways on different CPUs. The "fast-path" used a half-switch into the kernel by mapping off registers, allowing IPC to degenerate into register passing largely identical to a procedure call. As long as the message fit within the limitations -- 8 registers, no port identifiers, etc. -- it was faster than a traditional Unix trap. These limitations seem serious, but were in fact used for 80% of calls and 60% of returns (you often say "getDiskSector(integer value)" which could fit into the fast-path, and get back 2k of data which wouldn't).
Maury