Linux Runs All of the World’s Fastest Supercomputers
linuxfoundation.org
linuxfoundation.org
I know Linux fans don't see it that way.. young me thought Linux was indomitable and it was foolish for any company and group of people to not invest all their time and efforts into it. Older me knows better, that's not how corporations and groups of people work best. It leads to stagnation and apathy. It kills the fertile ground for ideas needed to make non-linear progression.
I don't know the solution to all this, but systems software is really going the way electronics engineering and circuit design have gone. I am sure executive types without engineering background would love to "outsource" it all to the perceived interchangeable cogs in "low cost regions".. but the rate of technological progress is about to drop of a cliff if my generation doesn't pick up the torch real soon now.
That said, it is somewhat premature as another commentator points out IBM's CNK. CNK is inventor's thinking.. the gall to ask "do we even need a general purpose OS?" regardless of what the market or eventual answer may be.
>>I know Linux fans don't see it that way..
I am not a Linux fan per se (as I type this from my Win10 desktop), but I legitimately don't understand what your comment is about. I've read it a few times and I'm still not exactly sure what you are complaining about.
How is Linux a monoculture, or this the endgame... or this stagnation?
I can't figure out what you are saying, sorry.
Or, just imagine microsoft won and you substitute windows for linux in this thought exercise. Wouldn’t that be a problem for innovation?
The TOP500 list does tend to be oriented to highly parallel systems because of the benchmark used. However, the different architectures like vector systems have largely gone by the wayside anyway. As CMOS scaling runs its course I expect we'll see more experimentation. Whether some of that requires a new operating system model as well remains to be seen.
There is no "winner" as far as I can tell in the grand scheme of things. We're talking about a set niche of computing. Win10 is still superior for a lot of GPU-clustering specific work; HBM2 memops via Vega architecture comes to mind as the Linux compute drivers are simply inadequate and the Windows tooling is significantly better.
Most companies, including very big ones with a long history of OS development are abandoning their efforts in favor of just using Linux. The perception is often that there is no innovation that can be done on the OS layer that will earn back its investment.
This leads to a lack of experimentation, tunnel vision, and in the long run stagnation.
I don't think the situation is nearly as bad as portrait here.
They can always use *BSD as a second choice too.
Additionally, where's the stagnation we are so worried about here? Any real world examples?
BeOS? Singularity? Plan 9? Hell, just read the last 20 years of OS research and realize that large swathes of it will never get used because anything that gets implemented has to be compatible with the way the Linux kernel does things. (By way of analogy, consider the x86 architecture, also a '70s era technology, and the massive contortions Intel had to go through to maintain compatibility over the years while adding new features and how all the newer alternative architectures died in the marketplace because x86 was too entrenched.)
Fuschia and, ironically, Windows are the last hopes for advancing OS architecture and that's really only possible because the semi-monopolists that created them have enough cash and engineering talent to push them until they take off.
We are talking about HPC computing and how Linux being the only viable choice leads to stagnation.
You are proposing those OSes advance the status quo (and I agree) but how is it that HPC computing is stagnated because Linux is the only viable option? If anything, HPC would be an awesome testbed for those experimental OSes and they still choose Linux.
Where's HPC stagnated because awesome ideas from other OSes can't be used?
Finally, are these other ideas so extremely awesome that they require a new OS from scratch (with all the uphill battle that it's to create it)? Can they be implemented on top of the Linux kernel? In which case, is Linux's success really such a bad thing?
I'm not saying this is necessarily a problem, but it risks becoming a problem.
It's clear we're seeing reduced investment in Linux competitors in exchange for Linux flavours from large players the past decade or so.
A 9 year old can install Linux and experiment and "innovate" all she wants. Linux's open source nature and the community is what powers its ongoing evolution, and that is encoded in its essence. I fail to see how the vein of Linux innovation on big iron is any close to being spent.
Which is absolutely wonderful, because Linux is licensed under GPL, so the best improvements end up in the main tree and everyone benefits.
Without taking sides in the debate, grsecurity/PaX patches are a counterexample.
It kind of prove Linux kernel is not mono-culture and natural selection and evolution are happening.
I believe the AOSP's kernel patches are also not mainlined - also proof Linux kernel are evolving very healthily.
Maybe it has to do with the Linux kernel being malleable enough to meet these specialized needs. My comment is a bit open ended, there's probably good reasons like the ideas I referenced had underlying issues. Microkernel security isn't really necessary on a supercomputer where anything running on it is more or less trusted code
It is not a monoculture, it's a common framework. Most farms have soil, and sure, some farms can operate without soil, but there is a lot of flexibility with soil. No two soils are alike, soil developers at each supersoil centre^W^Wfarm are free to derive and innovate as much as they'd like, but soil is a good place to start.
If you want a microkernel, port the Linux drivers to it and you're off to the races. If you want a monolithic kernel, there's basically nothing you're missing out on with Linux.
Not gloating – expressing pride in an open collective effort of hundreds of corporations and thousands upon thousands of individuals.
> monoculture
Only a bad thing if proprietary – witness how many Linux distros there are, consider the difference between Android and whatever runs on a supercomputer.
> I don't know the solution to all this
There's no problem, hence no need for a solution. You want competition? Then join any of the other free/open OS efforts and contribute your time and energy.
> Older me knows better
I don't think so. I prefer young you.
[http://quetzalcoatal.blogspot.com/2013/02/why-software-monoc... article), the first google result for "software monocultures bad" has a bunch of examples of bad effects from monocultures, even in open source software.
I am in Systems and Kernel development at least take a look at some commit data before you spew your unwarranted pessimism next time.
We are hitting that s curve plateau, and unless we have something like another world war (and even that may not be enough, as i do not get the impression we have anything sitting in the wing that is just lacking in funds to go massive) to open wide the purse strings i suspect there will be only a plateau from here on out.
I've sighed while removing code for systems like Irix, and Super-UX (which I thought was dead, but seems not to be). Their extended Unix resource management features were specifically useful for HPC-ish batch systems. The deficiencies of Linux are partially mitigated if you restrict the system to one job per node at the likely cost of wasted cores, but then -- like Bluegene -- you can ask what the kernel is there for. An example of the issues is lack of a sane way to manage and account memory for a job's process groups; please don't say that the cgroups mess provides that. No, OOM isn't what's wanted when you hit a cgroup limit, for instance.
Then we have no real solution for security in GNU/Linux HPC systems -- or other mainstream systems (even if they did have HPC hardware support).
Bluegene was interesting, but is now dead.
I just have an outsider perspective, but it seems that he still has to intervene every now and then to maintain a coherent architecture. Could anyone else fill this role, ie. be willing, able and empowered?
There is a possibility that Linux splits into multiple variants (server, supercomputer, desktop, embedded, ...), but as long as the licences are compatible I see no problem.
Just the other day there was a spat between a security guy and Torvalds, where the security guy wanted to push a patch set that would crash even the kernel if something in userspace made it do something "funny".
And said security guy have a whole bunch of cheerleaders in userspace (quite a few stepped onto social media to defend the patch set and berate Torvalds for being a potty mouthed bully), and perhaps even a few in kernel space.
As portrayed by tech-media Linus is the main man behind Linux, but for anyone who would even casually follow the LKML, "Linux" is merely Linus' Linux.
If Linus were to cease development, distributions would probably gravitate towards another Linux tree. There is no shortage of trees maintained by very qualified long-term developers who are trusted to keep everything in working order. They're just not as much in the general spotlight as Linus.
Even if different distributors would lean towards different trees, features and patches would be equalised by third parties and another main tree would establish itself as "Linux."
Except the mainstream desktop. Linux has not dominated the desktop, nor do I expect it ever will, at least not the desktop in its current form.
(I remember back in 1999 I would tell people that Linux would eventually rule the desktop, we just had to take a long view of at least 10 years. I can now say empirically that 10 years is not a long enough view.)
I'm not saying it's perfect, but in 2017 and being a developer, I'd take a serious look how nice of an OS it actually is.
I worry more about constant willfull breakage with legacy within said desktop, compared how MS still supports Win32 (not so dissimilar from Torvalds "do not break userspace" kernel policy).
That said, whenever Linux have appeared to gain a toehold (like with netbooks), MS have dropped lisencing prices and employed rebate programs to make Windows productes the "cheap" option.
The HPL number is acknowledged as bad, even by the standard of single performance numbers. It basically reflects the number and width of the vector units and use of a good BLAS. https://hpgmg.org/ seems a better choice if you must have a single HPC benchmark number. (HPCG, touted to replace HPL, depends on how you re-write it, which you must to avoid essentially STREAM results.)
It's also worth pointing out that much of the capacity there isn't even on an OS kernel, but is off-loaded, and I wonder what multiple entries for "service provider x" actually are as "systems".
Edit: Just for clarification, I didn't mean to suggest that the linux foundation did lobby. I just wonder if they were thinking about it, and what their decision was.