The only time I faced the need for Linux box is trying a demo project from OpenAI, which did not use the features it required Linux for on a single machine anyway.
Doing something custom on top of the GPU is also not much different on Windows, than Linux. CUDA is basically the same. OpenCL and Vulkan are available too.
I'd like to hear perspective of a person, who actually does ML specifically on Linux for some reason.
Also windows not having a build in C compiler makes you dependent on the horribly convoluted Visual Basic stack that seems to have a lot of dependencies for some python ML libraries. Docker makes it a lot better to run and I almost always deploy in a docker container because the ML modules I deliver are often interacted with as a black box with a REST API on top.
You might be right about C compiler. But something itched when you mentioned Docker. Could getting Windows SDK installed be harder, than installing Docker?
That being said I had to avoid installing VS 2019 for quite awhile because Node.js native module build chain couldn't work with it. There are complexities
Even popular libraries like zeromq don't support namedpipes on windows because of how complicated they are and how different to everywhere else.
Just determining what visual studio version is installed seems to trip up projects all the time.
My main reason: It is the most convenient way to have Unix tools (grep/sort/cut/sed/less/...) and bash available. Cygwin always was a pain, MinGW / GitBash felt much better, but ultimately WSL just feels best.
These tools are incredibly valuable to my workflow. Sure, stuff like pandas can be nice for small datasets, and some data sits in some DB/Kafka/distributed system. But there have been countless cases where unix tools allowed me to take xxGB zpfiles of text and do basic examination or even build baseline models within a few hours.
Sure, there always are alternatives to use these tools and there are many equivalents. But I would always prefer WSL + conda for Linux to a typical "Windows Conda" installation with that weird GUI and the need to install so many different applications to even just look into the first or last few lines of a huge textfile.
EDIT: That said, of course I can/could always just run a juypter notebook under windows using windows cuda + GPU and share files with a WSL bash where I do my modifications. But again, everything within the same systems just feels better (ipython shell magic, no worries about if paths to the same file are really identical, etc) and while this is by no means a game-changer, it is just nicer that way.
https://github.com/pytorch/pytorch/issues/37790
weird
https://github.com/pytorch/pytorch/issues/32575
bug
https://github.com/pytorch/pytorch/issues/25301
I don't have to spend time explaining or justifying or isolating my setup.
Conversely though, my work is itself off the beaten path enough that I'm likely to run into weird bugs. If I was pushing images through a CNN, that'd be well-trodden enough on every platform that I'd be a lot less fussed about which particular platform I use.
Not sure why, but I suspect that Valve has a lot to do with it.
Conversely, if you run Windows, it's rare that you need to work hard to run a game.
But there is another way there - running Windows in a VM with GPU passthrough - works beautifully in my case.
Like the previous response what does this buy me as compared to developing and running natively in windows - as there are native compliers that support cuda etal on windows.
Check out https://old.reddit.com/r/VFIO/
I guess this will be the standard until we can have nicer graphics drivers for Linux.
That's not the usual situation for people that are trying to use this scheme.
If you have a spare GPU, a VM with PCI passthrough does an even better job, except for some anti-cheat software that artificially discriminates against this setup.
In theory it ought to be possible to switch a single GPU to/from a VM without a reboot. In practice I have no idea how huge a refactoring to the Linux graphics stack that'd require.
This has been true for 16 bit games since long before Windows 10. Ages ago one of my favourite games stopped working on Windows, but Wine had no problems with it. So my impression has always been that Wine is excellent for really old games, but slightly more recent games, it could already be very hit and miss.
> "If you have a spare GPU, a VM with PCI passthrough does an even better job, except for some anti-cheat software that artificially discriminates against this setup."
Doesn't every CPU these days have onboard graphics? My Thinkpad X1E should support hybrid graphics, so it'd be nice if I could give the GPU to a VM and have the desktop use the CPU graphics.
But if a Windows VM does a better job, that means Wine doesn't yet do as good a job as Windows. Though it's certainly true that Steam support for Linux is growing. But I don't think every Steam game already works on Linux.
On laptops, pretty much. On Intel desktops, yes, aside from Xeons. On AMD desktops, only some lower end Ryzens have "G" models.
Official Proton "support" is limited, because it requires certification by Valve and/or the game developers that the game runs well (the equivalent of a "native" rating on winedb/protondb), but if you're willing to go down to "gold" levels of support it still runs 70-80% of all Steam windows games.
Don't bother attempting GPU passthrough on any laptop with an AMD CPU (eg Ryzen 2700U) and Radeon GPU (eg RX 560X).
It turns out the GPU passthrough needs a dump of the Radeon BIOS provided as a file, but no-one can dump the BIOS of discrete AMD laptop GPUs. :( :( :(
Note the complete lack of RX 560X BIOS's here:
I've got this working today. I do it through swapping the nvidia driver for the vfio-pci driver (and back again if required). The slight annoyance is that you may need to restart X11 (for me this is not an issue).
I wrote about this some years ago: https://me.m01.eu/blog/2016/05/pci-passthrough-vm-monitor-se...
(My gaming desktop, of course, runs Windows.)
On laptops, however, I continue to find myself jumping through idiotic hoops to use Linux. The driver support is always just-barely-good-enough. Maybe it's audio, maybe it's graphics, maybe it's power management. Maybe it's something involving networking-after-power-management or some crap involving "don't unplug your headphones while the lid is down". On my current work laptop, a beautiful Thinkpad Carbon X1, I can't get the power management to work properly, so I just have to accept that there's no hibernate. I'm constantly forgetting to shut down, put it in my backpack, and then pull it out drained. What a pain in the ass. Could someone fix this problem, probably. Can I? Not in the dozens of hours I've put into it. I hate doing IT, I hate it I hate it I hate it.
However much Macs make me want to vomit in my mouth, I can see the appeal. The drivers work at least 90% as well as Windows drivers, and the UX is at least half as good as a lightly-tuned Linux machine. "Jack of all trades, master of none, is oftentimes better than master of one"
Anyway, before this lockdown ends, I'm upgrading my laptop distro to this new distro I've heard of out of Redmond, I think it's called "Windows".
Not to mention that there are quite a few development environments for more obscure platforms that still only exist for Windows.
Overall, since most development time is spent in an IDE, the OS is really of little relevance to software development. Sure, some people insist in using command line tools, and that is unlikely to be pleasant on Windows, but a lot of other developers don't, and we couldn't care whether we're running our Emacs on Linux or Windows or Genera or whatever.
However, if I'm using IntelliJ or Emacs and Firefox, I don't really need to care what OS is running underneath too much.
Edit: of course, Linux and Mac are available for devs that prefer them. It's still much easier for IT to manage 7000 Windows desktops and a couple hundred Linux ones than it would be to manage 7000 Linux desktops.
- Could you push a patch to Linux systems and have it install at the user's convenience (with some end date)?
- Can you do that in waves without manually configuring things?
- Can you remotely wipe a system if required?
- Is there any popular anti-virus software for Linux, to protect company files in home folders from user mistakes?
- Can you help users install some software without giving them full access, but also without requiring IT intervention for every installation?
Why? I have never seen Windows being managed entirely hands-off whereas Linux just works.
Where does the complexity on Linux come from that makes managing them more difficult?
Then there's the question of pushing an update to all managed computers. Maybe it's not a package update, but you want to change some SELinux policy for all users, or update some DNS server or the default search domain and so on.
Never mind the question of how you can instruct one of those Linux computers to delete all data it holds whenever it next connects to the internet (to handle the case of a stolen company laptop).
There are so many things that you need in an enterprise setting that have common (though probably quite expensive) tools available for Windows. Maybe some of these exist for Linux as well (I would expect RedHat to have some), but I'm not sure. Linux admin is usually reserved for servers much more than desktop computers.
Interestingly, apples have to be compared to oranges. On Linux, it is easy to identify the programs that are using a library. Thus it is easy to restart just the services that are patched. In general, things can be scripted so there are no tools available. But this requires somebody who understands the system. From a business perspective, this might be more expensive, or not, if the tools are expensive.
I use a Windows laptop at the company I currently work for, because everything is locked down and I wouldn't be able to get my own laptop connected to the network. (Or so I thought; I co-worker managed to use the Windows laptop as a bridge to his own Macbook.)
Now you're right that as long as I stay in the IDE, it's not so bad. But every once in a while I need to do something outside the IDE, and I immediately get slapped in the face by how stupid some things are. And because it's an enterprise environment, some things are even worse than usual; opening a folder, or saving something, can be unreasonably slow because either it's a network drive or it needs to be checked for viruses and malware while I'm trying to use it. Or for some other reason. I don' t know, I just experience the extreme slowness.
Also, on top of the old terrible DOS shell, there's now also a Power Shell that's supposed to be better. It apparently has some powerful features I don't really grasp, but it's still not remotely as good as bash. And sometimes the command line really is unavoidable.
But the real pain is at home. When I activate Windows 10 on a new machine, I need to create a Microsoft account. I don't want one, but it takes serious determination to avoid it, because behind every message is another trick to sucker you into an MS account. When you finally do manage to create a local account, you're immediately expected to compromise your security with 3 insecurity questions, and no way to avoid it as far as I can tell. Previous versions of Windows did not have this stupidity.
Also, somehow Windows keeps losing my mic, speakers or camera. Once I've found the right troubleshooter, it immediately figures out how to fix it, which is great, but it also keeps losing them again. And finding the right troubleshooter takes a couple of steps and a bit of searching. I feel like I need to pin several relevant troubleshooters to the taskbar.
And then there's the total lack of access control. To install anything, you need to be admin. I gave my son a restricted account, but he can't do anything with it. I'd like to be able to create an account that can instal games, but can't compromise the system. No such option in Windows. If you can do anything, you can do everything. Unless you're in an enterprise environment, in which case you often still can't do anything. So I guess more detailed access control does exist, but only for enterprise users or something.
> And then there's the total lack of access control. To install anything, you need to be admin. I gave my son a restricted account, but he can't do anything with it. I'd like to be able to create an account that can instal games, but can't compromise the system. No such option in Windows. If you can do anything, you can do everything. Unless you're in an enterprise environment, in which case you often still can't do anything. So I guess more detailed access control does exist, but only for enterprise users or something.
Here I never understand this point. You can't do anything on a Linux system if you don't have sudo access - it's not like apt or yum have any special magic to allow non-admin users to install stuff. And if you can install software on a system, you can already do anything else. Especially Games, which install drivers to perform DRM and anti-cheat bull.
Now, if you want to look into it and waste quite a bit of time, Windows does allow you to configure access control at a very fine-grained level for access to non-system folders. But as long as the installers want to install things in system folders, there really isn't any solution.
aka wine / proton
How do you configure compiler to use libraries downloaded from vcpkg? CMake? Something else?
I do not think it's purely windows to blame here though. It's only quite recently that NVIDIA started fixing their documentation and instructions on getting all the right CUDA CuDNN stuff running properly on a system.
Imagine if you could run AI/ML apps and tools that are coded to take advantage of DirectML on Windows and/or atop DirectML via WSL.
Now you can run the tools you want and need in whichever environment you like ... on any (capable) GPU you like: You don't have to buy a particular vendor's GPU to run your code.
If you're old like me and remember the dark ol' days when games shipped with specific drivers for (early) GPU cards/chips, but failed to run at all if you didn't have one of the supported cards, you'll understand why this is a big deal.
Maybe I'm not that old, but I'm old enough to remember the days when microsoft was intentionally degrading opengl performance on windows ;).
Those days sucked. Bigtime. If we can avoid doing the same mistakes for machine learning then we should.
Which is still nonsense, since this only affected the OGL driver shipped by Microsoft. In contrast to truly bad actors like Apple, OEM were free to ship their own OGL drivers from day 1.
So sorry mate, but I have to call BS on that one.
Don't you think the effort to achieve this would be absolutely massive? I don't know what kind of resources are thrown on this project, but I'd estimate minimum to be 3 dev teams for 2 years just to get a few variations of ResNet to work "as is". And that's just for regular models, that don't require quantization or (auto-)mixed precision for training.
Isn't the linked post saying you have to be running on Windows though? It seems like it would make way more sense to either port directX to Linux, or ditch directX and put those resources into supporting Vulcan.
o-O nvidia-docker does not even support Windows.
I think the only thing you need to know is which CUDA version your cuDNN requires, and it was quite clearly stated on the download page. Also the same on Linux. For nvidia-docker you used to need a specific driver version.