"Linus Benedict Torvalds
In article <1992Jan29.2...@klaava.Helsinki.FI> I wrote:
>Well, with a subject like this, I'm afraid I'll have to reply.
And reply I did, with complete abandon, and no thought for good taste and netiquette. Apologies to ast, and thanks to John Nall for a friendy "that's not how it's done"-letter. I over-reacted, and am now composing a (much less acerbic) personal letter to ast. Hope nobody was turned away from linux due to it being (a) possibly obsolete (I still think
that's not the case, although some of the criticisms are valid) and (b) written by a hothead :-)
Linus "my first, and hopefully last flamefest" Torvalds"
This HN comment posted from a Linux machine in 2015. :)
I find it interesting that this post can be taken two ways.
1. It's clear that Linux 'won' in the marketplace so we can all laugh at how wrong this guy was and the curiosity of these 'microkernel' things and that portability stuff. Lols all around.
2. We've reached a point where ideas are gaining ground about immutable infrastructure, people are talking more about things with similarities to microkernels called unikernels [1] (and where they might be used [2]). Linux isn't going anywhere but these new approaches have value and were being discussed as long ago as 1992. Of course, the author got things wrong but that's par for course. It's more interesting to see ideas that are resurfacing. Incidentally, ARM was a RISC chip and now dominates Intel on mobile devices.
I prefer the second approach. So to anyone poking fun at the author, please consider that maybe this is one aspect of living in the future [3], albeit much further than than most (market timing is always underrated and academics tend to think further ahead than most people). We can also remind ourselves that 'Better' is a tricky and subjective thing to define (cf VHS tapes vs Betamax).
So we now live in a world where both monolithic and "microkernel" (hybrid) systems are commonplace - there's a bit shy of a billion devices running linux (android) and no shortage of windows and osx hosts around. What can we learn from this?
Well, it doesn't really seem to matter whether you use a monolithic or microkernel design. They both work well enough. Driver development is hard for reasons that have got nothing to do with this, and doesn't appear to change much between the platforms. Success/popularity of operating systems appears to be determined by other factors.
I'm not sure why people feel this is still worth debating. All the evidence points to it being completely unimportant.
Depends on how you look at it. Darwin is based on Mach with some monolithic BSD stuff linked in, I remember reading that Windows was fairly microkernelish but they've been pushing more things to kernel space.
Arguable. There might be more instances of QNX and L4 running than of Linux. This is pretty much impossible to measure. Remember that Android smartphones have a second OS underneath or beside the Linux kernel.
Apple and Microsoft use a hybrid approach, so they just do not participate in the debate.
> OKL4 shipments exceeded 1.5 billion in early 2012, mostly on Qualcomm wireless modem chips. Other deployments include automotive infotainment systems.
I would very much like to laugh at Andy Tanenbaum and we all should. He deserves to be ridiculed because he was being an a-hole. He was putting down another programmer based on a theoretical mumbo jumbo when he could have proven his theories by actual programming.
I think it is obvious that any scientist that can prove his theories in a practical way must do that and should not flame and criticize another. In areas where theories are harder to prove like astrophysics, social science, economics, etc. I guess people can debate and flame. We are all used to economist flame wars by now.
But in areas where theories are practically provable, such criticisms are useless and idiotic. If Tanenbaum thought microkernels were so great he should have made a usable microkernel OS.
This also shows a very negative practice in the software industry which should be discouraged. And this is the criticism of the thing that is working now in favor of some new shiny project that is still in development but it will be all so great when it comes out. It is a very easy pitfall to get into, because the thing working now usually has some problems and disadvantages while the shiny new thing that does not exist yet is usually perfect in every way. But in practical matters, one usually does better going with the thing working now. But worst of all, this practice of worshiping the nonexistent solution is usually used to strike down newcomers and disruptors.
By the way, unikernels have absolutely nothing to do with microkernels. So, no the day of the microkernel has not arrived and there is no indication that it will come any time soon. Maybe there will be a practical working microkernel based os some day. I have no idea what will happen in a thousand years or ten thousand years. But if that happens it will be more of coincidence than evidence that Andy Tanenbaum is some kind of genius living in the future. If you have long enough time frame a lot of predictions about the future will eventually happen in one way or another.
Regardless, Linus never disputed the fact that microkernels are theoretically better. He just knew how impractical they were because he actually buckled down and did the work to make a useful OS instead of bloviating on theory.
>We've reached a point where ideas are gaining ground about immutable infrastructure, people are talking more about things with similarities to microkernels called unikernels [1] (and where they might be used [2]).
Not only this. Amoeba, Tanenbaum's research OS is a distributed operating system, meaning that many devices could be used as a single system. This would be incredibly useful in our current age where everyone has a laptop, a smartphone, a tablet, and lots of other devices that need to be synched.
Usually distributed OS are designed for a scenario where every device is (nearly) always on. Like a supercomputer or a LAN of desktops. Mobile devices want to be mostly off to save energy.
it depends a lot on what you mean by won and which marketplace you are talking about.
as a desktop os linux is essentially non existent with an ever shrinking 1.5% of the market share or so as more and more regular joe consumers get machines with windows or os x preinstalled at a much faster rate than expert-level geeks are installing linux.
as a software developer targetting the mainstream it is rare that targetting linux is a profitable venture, because nearly no customers use it and those that do don't spend money on software in the same way as some one living in the Microsoft or Apple ecosystems. OS X also has a tiny market share (around 4% iirc) but people who are willing to pay over the odds to have a pretty workstation seem also not afraid of spending money on software.
as a server platform Linux dominates despite the best efforts of MS and Apple to try and market slightly uglier versions their consumer level desktop OS as a server product. its easier to use remotely, and can easily be configured to not rely on heavy UI features - there are even flavours that cater better to business philosophies on 'stability' (i use that term loosely) like RHEL and CentOS - also stability and open source go hand in hand, because you can fix bugs, or pay people to fix bugs allowing you to throw money at problems to fix them in a way which is impossible with the mainstream OSs
as a software developer it makes sense to target linux first for your big expensive server product. if you don't then you are cutting out your customers...
i don't think either of these situations has much to do with the underlying technology. its much more to do with the OS and what it provides than how it is built to achieve those aims.
Chromebooks (which use linux) accounted for over 5% of laptop shipments in Q3 of last year. They also grew at an astounding rate of 67% quarter over quarter. I have not seen data for Q4 of last year yet, but considering the chromebooks' growth rates it will probably be even higher.
So, no desktop linux market share is not non-existent and it is definitely not shrinking.
Given the way Google is steering Android these days i am not sure i want it on a "desktop" any time soon.
Back during early 3.x it had a spark of potential. But since then Google has done quite a bit to lock it down so that it is basically a media consumption terminal.
> But since then Google has done quite a bit to lock it down so that it is basically a media consumption terminal.
It remains possible to choose to install, or not to install Google play Services and Google's app suite on aftermarket Android installations, and still be running Android, and running most 3rd party Android apps.
Any app developers who wants to run on Amazon, Jolla, Blackberry, CyanogenMod, and on the embedded Android in cars and appliances, and all the Android variants in China is going to have to take care to make use of Google Play Services optional, and to correctly use Intent filtering to make of of and respond to high level IPC messages to and from other apps, no matter if they are the "official" Google apps or not.
So, while Google could be more open without risking much if any advantage, Android hasn't been locked-down. Google has little chance of doing an "embrace and extend" on Android ISVs. You still have a wide range of choices.
This comment is exactly the opposite of reality. 3.x was the most locked down Android ever. Its source was never released. Since then, AOSP has gotten strictly more functional with every release. The proprietary Google apps built on top have gotten wider distribution on the devices that have Android installed, but that doesn't make AOSP itself more locked down.
From a user perspective, Android has gotten more locked down. This in that while previously one could access the file system directly, now on has to go via a sub-system that can deny or hide access arbitrarily.
That’s the nature of any consumer-oriented product, though. Unless someone starts making a phone for hackers, they’re all going to be designed for media consumption.
With that said, you’re right about the locking down. It’s a good thing that we’ve CyanogenMod and Replicant.
There are few "big expensive server products" now - people are running SAAS, or open source products that may or may not have expensive support contracts. If you run SAAS you can run anything you like, look at Whatsapp with their Erlang on FreeBSD stack.
> Linux 'won' in the marketplace so we can all laugh at how wrong this guy was
Oh, it has been a long time since 1992 and these posts by Tannenbaum are pretty famous (actually, I'm surprised it still appears on HN). So we did laugh already. Later we suddenly stopped laughing and it was more like your "second approach". But that also was quite a while ago. Next wave of philosophical though in techno society was reflecting on all these "worse in better" stuff, "MIT vs Berkley", "Lisp is older that C", "ML is 40 years old". Because, really, it's all about the same idea. And posts about "Cathedral and bazaar" and how dot com bubble destroyed the notion of sound architecture is pretty much about the same thing.
So for me it isn't anymore about any new "lessons" to learn from that. Just a painful reminder on "sooner or later we'll have to throw all these tremendous work away and start anew… shall we?".
Tanenbaum said that Torvalds would not get a good grade on his course, epic.
"I still maintain the point that designing a monolithic kernel in 1991 is
a fundamental error. Be thankful you are not my student. You would not
get a high grade for such a design :-)"
Perfect display of how universities are good at judging people about how well they know how to "play the game" (and usually that involves conforming to whatever frame of mind your professor thinks is right).
Well, a microkernel is harder to implement than a monolithic one and from a theoretical standpoint is a better design, IMO it deserves a higher grade in an operating system design course. Let's not mythify the 1992 Linux kernel, which probably wasn't _that_ good.
A microkernel is much easier to implement than a monolithic one. Been there, done that. The hard part is to get the messaging done with zero overhead (no copying), but using paging for the mbufs you can get quite far with that.
A microkernel is certainly easier to build than a monolithic kernel, because it does much less. You want to compare a feature-equivalent system, so microkernel + some userland drivers etc vs monolithic kernel.
The monolithic approach is easier because you just call a function in another component or read its data structures. In a microkernel, you need to design interfaces and protocols first.
No, the monolithic approach is definitely not easier for a simple reason: there is hardly any kernel level debugging going on in a microkernel based system because the kernel is small. Everything else is userland, so you can test your new driver as just another user process. This takes the sting out of a very large chunk of frustrating debugging and gives you access to all the userland tools to home in on the bug.
And even better: a crash of your driver does not take down the whole system. So you can just keep on working.
Designing protocols and interfaces is roughly the same in either case, after all you could settle on a very simple set of messages for most interface problems, with open, close, fcntl, read and write you would be able to do the majority of interface tasks.
If, as Tanenbaum did, you look at it from a computer science perspective, Linux wasn't good, indeed. It was 'just' a Unix clone.
From a software engineering perspective, it left a lot to be desired, too. It basically only supported what was on Linus desk.
One could say the things Linux had going for it were a) a GPL license, b) being small, c) arriving at just the right time, and d) Linus willing to accept external input.
Whether a) is an advantage of course depends on personal preference. b) helped with d) because it made it feasible for many to download the thing; c) was luck; d) IMO was the most influential factor.
In theory you can postulate the performance impact is worth. In practice the is much more in it. That is the difference between computer science and software engineering.
IPC overhead was primarily an issue with the Mach kernel in particular (due to it performing elaborate checks on message ports) that later ended up unfairly stigmatizing the entire microkernel design in general.
Contemporary microkernels like L4 are much, much faster.
There seems to be a response by "the" (please rise) Ken Thompson:
"viewpoint may be largely unrelated to its usefulness. Many if not most of the software we use is probably obsolete according to the latest design criteria. Most users could probably care less if the internals of the operating system they use is obsolete. They are rightly more interested in its performance and capabilities at the user level.
"I would generally agree that microkernels are probably the wave of the future. However, it is in my opinion easier to implement a monolithic kernel. It is also easier for it to turn into a mess in a hurry as it is modified."
Actually, I always enjoy reading this whenever it comes up. I was just about to start a university computing science degree course at this time and was reading Tanenbaum in preparation. This was a debate that was alive for years throughout the computing community I was part of.
In the end reality came down to what always wins in computing (and in life too).
Running code beats design. Design is always 'better'.
There's a variation to that. Unless your company is a very successful multinational chances are that a stranger picked at random has not heard about your product. Makes you think a little longer about terms such as 'market saturation', most products don't even go near there.
The problem with this whole debate is that Linus was 'more wrong' than Tanenbaum and Tanenbaum made plenty of mistakes with Minix which also make it more of a macrokernel than a true microkernel.
A true microkernel does one thing and one thing only: pass messages.
Now that's an ideal and in the real world you don't get to have your ideals realized so rather than to be able to realize this spherical cow you're going to have add in a few more system calls to make it work but you'll end up with something a lot closer to plan9 or QnX and compared to those Linux is very very old hat indeed, it's basically a re-run of the 70's state of the art with a a whole pile of modern day hardware drivers and other goodies thrown in.
Tanenbaums biggest mistake was to try to monetize Minix through Prentice-Hall, if he'd just tossed it out there it would have picked up steam a lot quicker, but likely he too had bills to pay and his expenses at the time were probably a lot larger than Linuses, and so history was made.
So, Linus was obsolete, but so was Minix and the future as we could have had it is still waiting to happen. And when it does you'll finally appreciate just how obsolete Linux was back in 92, and how much more obsolete it is today.
Until then it's like democracy: not perfect but the best we've got (without shelling out lots of license fees for something better).
> Tanenbaums biggest mistake was to try to monetize Minix through Prentice-Hall,
You make it sound like his capitalist motives worked against him.
Writing and publishing educational material was part of his job. Prentice hall had been publishing his books since the 1970's. I'm sure it seemed like the obvious way to publish his educational material - especially since minix wasn't a standalone thing, but it came with a book. Perhaps he even had an exclusive contract with PH. It probably didn't even occur to him that there was another way to publish the software.
Also, in 1991, making software available for download wasn't as straight forward as it is today. There was no www, and very few Europeans had access to the internet. I know for me a book with a disk would have been the only reasonable way to get it.
In 1997 in the suburbs of Chicago, a book with a CD was the only reasonable way for me to get linux... By that time I "had the internet" but it was 33.6Kbps dial up, beginner resource pages were lacking (so a book to teach me was nice), and downloading a full linux distro was going take to way too long (so CD was the way to go). Particularly as a 17 year old without a credit card to order CDs online from one of the early, untrusted, online merchants.
Cost was not the only problem with Minix, although it was pretty hefty (it would have been about $180 in todays dollars for me in the UK). For students (as both I and Linus were) that was a lot of money.
AST was firmly of the belief in Minix as being educational code, and hence only accepted patches along those lines. I don't remember if there was copyright assignment too, but there must have been something similar. This was at odds with the people who wanted to use Minix as their operating system.
For example the Minix of the time only supported 16 bit mode on x86 which meant no memory protection and only using about half a megabyte of memory on multi-megabyte machines. Someone produced fairly large patches that let it work in 32 bit mode, but they were never distributed as part of Minix because they didn't further the education goal (more code, harder to read and understand etc).
Another example was that the filesystem code was single threaded and only handled one request at a time. This meant easy to read and understand code, but dismal performance. Again this hindered those trying to use Minix versus those learning from it. This was such an issue that the original Linux announcement has this line:
PS. Yes – it’s free of any minix code, and it has a multi-threaded fs.
Essentially Minix wasn't seriously usable, cost a lot, was not open source/community oriented, and was not progressing towards being usable. That is why so many of us jumped on Linux because it immediately worked, didn't cost anything, was 32 bit, had a rapid pace of development, and was trying to be usable.
On the Hurd side, there never was anything you could actually download and use. It was also developed very strongly in the cathedral style as many GNU projects were. Of course the promises were that "soon" there would be something. But Linux was there now and accepted patches without copyright assignment, and had rapid releases.
Hurd simply never was ready and/or welcoming outsiders that might have pulled it in a direction where the anointed did not want it to go (such as, actually releasing something).
Hurd and Minix both missed critical windows of opportunity and ended up so far behind Linux that they might as well not have bothered. For the longest time the only thing that kept new people flooding into Minix was Tanenbaums courses at VU.
Hurd was a research project as much as a working kernel.
I Think the project restarted 3 times in 10 years...
Linux may have been crude etc, but it was here, it was working, and it was free (in both senses of the word).
Frankly i see the lack of Linux on the desktop less about the quality of Linux, and more about MS abusing its market position and a massive failure to regulate.
"and more about MS abusing its market position and a massive failure to regulate." thats a pretty vague statement.
IMO its simply the fact that Windows/OSX have set a high bar of out-of-the-box usability and shine that GNU/Linux simply hasn't reached. Not that this is a bad thing, it just might be worth accepting that we can't compete on desktop.
Back during the netbook froth, they put XP on life support using a highly focused license that ended up defining what a netbook was (in essence a small, underpowered, laptop).
This while the early netbooks from Asus and Acer ran Linux installs that can be likened to ChromeOS, with a UI somewhat like Windows 8 Metro.
What they did may very well be considered market position abuse.
Hurd failed mostly due to managerial incompetence. RMS was insistent on sticking with a Mach base, which at the time was both unfamiliar to his team and encumbered in licensing limbo. Thomas Bushnell, the former Hurd project leader, has said that if his plan to adopt a 4.4BSD-Lite kernel had been greenlit, then Hurd might have succeeded and Linux could have been a footnote.
Of course, if the BSD kernel plan went through, then Hurd would be quite different architecturally today. I'm not sure how versatile it actually was, though - perhaps a multiserver design still might have been possible.
Tanenbaum admitted at EuroBSDCon last year[1] that he was wrong about BSD and he should have turned Minix into a BSD, which he now has, with Minix3 borrowing NetBSD userspace, build system and so on.
Something to meditate on any time there is an "Appeal to Expert" - particularly on an issue that is forward looking.
Also, I love how he was so confident with regards to
While I could go into a long story here about the relative merits of the two designs, suffice it to say that among the people who actually design operating systems, the debate is essentially over.
The funny thing is that I would mostly agree with him. The important detail is "Among the people who actually design operating systems." Linus would agree that Linux is not designed. It evolved. Very few OSs are designed these days.
The QNX microkernel is quite successful. L4 also seems to do well. The BSD Mach microkernel is still around. Minix is still around, although its reliability approach (similar to Erlang, let processes/drivers fail and restart) does not gain mindshare. Windows is considered a hybrid since NT. Likewise XNU/Darwin is Apple's hybrid.
Maybe the current state is more like: Neither micro nor monolithic won, but we know the tradeoffs now and an OS designer can decide per feature. For most applications the difference does not matter, though. Most applications build on a higher level platform (JVM, iOS, Browser, etc).
Also this (later in the thread), he was pretty wrong..
"Making software free, but only for folks with enough money
to buy first class hardware (x86) is an interesting concept.
Of course 5 years from now that will be different, but 5 years from now
everyone will be running free GNU on their 200 MIPS, 64M SPARCstation-5."
How feasible is it to radically refactor an operating system? Would it be possible to gradually incorporate all the insights of operating system research into Linux/BSD/$otherProductionReadyOS or is it always necessary to start again from scratch?
Minix3 is to a fairly large extent a refactor of NetBSD to have a microkernel base; osx kind of did a similar thing with freebsd. Also there are ways of running old OSs on top eg Linux on Genode.
NetBSD is a good starting point as the drivers are portable (via rump kernel; eg Genode uses this for their filesystem drivers) so you can reuse them in another OS as a starting point, plus it is reasonably simple as OSs go, plus the BSD license is friendly.
Through refactoring Linux has been transformed to be a real time kernel using the linux-rt patch set. The patch set is aggressive and big and only slowly making way upstream.
Still it proves that yes, refactoring can work. Real-time is one of those super-duper invasive features. So if it works for real-time we should be able to refactor in other features.
Looking back it's amazing how much traction the gnu userland tools gained and how little (zero?) their kernel gained.
I'm not going to call it a wasted effort, because I appreciate there being other "Unix"-systems around in case Linux ever goes bad, but at this point it seems to be a whole lot of work invested for nothing.
Maybe things will change 5 years down the road when people abandon systemd and DMD had matured, and the best way to run it is on Hurd, but I'm not putting any money on that bet ;)
There are the BSDs as other Unix systems around; they of course were not at the point when this was written as the lawsuits were unresolved. NetBSD started in 1993, after this. So we have plenty of backup without having to fix Hurd, as well as the L4 based microkernels that are much more mature and interesting than Hurd.
Is it still possible to asscess comp.os.minix old archives through NNTP somewhere? Google Groups have the old emails but I can't find a way to read them in a local email client.
"Linus Benedict Torvalds In article <1992Jan29.2...@klaava.Helsinki.FI> I wrote: >Well, with a subject like this, I'm afraid I'll have to reply.
And reply I did, with complete abandon, and no thought for good taste and netiquette. Apologies to ast, and thanks to John Nall for a friendy "that's not how it's done"-letter. I over-reacted, and am now composing a (much less acerbic) personal letter to ast. Hope nobody was turned away from linux due to it being (a) possibly obsolete (I still think that's not the case, although some of the criticisms are valid) and (b) written by a hothead :-)
Linus "my first, and hopefully last flamefest" Torvalds"
This HN comment posted from a Linux machine in 2015. :)