Slashdot Mirror
Ask Slashdot: Can Linux Run a GPU-Computing Application Written For Windows?
dryriver writes:
I have been told that Linux can run Windows software using Wine or perhaps a VM. What happens if that Windows software is a GPU-computing application -- accessing the GPU through HLSL/GLSL/CUDA/OpenCL or similar interfaces?
Can Wine or other solutions run that software at a decent speed under Linux? Or is GPU-computing software written for the Windows platform unsuitable for use -- emulated or otherwise -- under Linux?
This sounds like one of those cases where there's a theoretical answer and then your own real-world experiences. So leave your best answers in the comments. Can Linux run a GPU-computing application that's written for Windows?
Maybe.
Just cruising through this digital world at 33 1/3 rpm...
PCI express passthrough is the term to search for.
A good writeup:
https://www.reddit.com/r/pcmasterrace/comments/2z0evz/gpu_passthrough_or_how_to_play_any_game_at_near/
It used to be there was no Linux Folding@Home application. But there was a way to run the Windows CUDA application on Linux, and I did for a couple of years.
Here's a guide for that: http://www.overclockers.com/de...
(T>t && O(n)--) == sqrt(666)
Even if it runs under something like WINE, how fast does it run?
I would assume that if you need something that uses CUDA, performance is important.
The real "Libtards" are the Libertarians!
Lots of Linux and crossplatform applications have been using GPU regardless of the OS for a while now. Blender is a good example. Sometimes all it takes is the right drivers. Please tell me you're not trying to run Maya or something like that when you got Blender? :(
Run Windows. I never have used wine. Linus himself said if you need to run win32 then run host OS. Linus used XP for PowerPoint.
How would you feel if I asked you if I could run Ubuntu on Windows 10? I could but why? It's a geek toy. At work your job is to put ideology aside and to make your bosses happy by keeping it up.
Will running Windows Server really be that horrible for 1 server?
But if I did have to run Linux I could try KMS with GPU pass thru but you risk downsides on now 2 platforms and bugs for now KMs, Linux, and windows and a performance hit. That won't guarantee higher uptime.
http://saveie6.com/
If it is GPU bound, then wine is perfectly fast enough for acting as a 'windows i/o emulation layer' to pump the data to the cards (barring any serious regressions, wine is at worse usually 2-3x as slow as windows, due to the extra API overhead. But in some cases is faster due to filesystem, memory or i/o subsystem, etc.) In that case, so long as the wine API can pass through all the GPU programs needed and pull the results back to the windows apps, it should work exactly the same as on Windows, but with the benefit of a few hundred to few thousand dollars worth of licensing that can be put towards GPUs instead (assuming of course the apps in question are not proprietary or have DRM which restricts them to running on authentic windows systems only. In that case, you might instead consider a VM+IOMMU platform using either xen or qemu/kvm with PCIe passthrough to run a real copy of windows and your GPU app, subject to licensing restrictions.)
Just like I wouldn't go the other way and do a Linux computation program on Windows. Thing is if you are talking about something that needs the GPU for acceleration, you are talking about something doing some heavy hitting math. The reason you bother with a GPU is because whatever you are doing is too slow just running on a CPU. Otherwise why bother? It is more work and restricts systems that can run it.
Well any emulation/virtualization/whatever will slow things down. How much varies, but even in the best case scenario, there's a performance penalty. Why pay that? Just run it in the OS it wants to run in on the iron. That'll get you the best performance.
Not only that but generally speaking technical and engineering software is not the best written, most stable stuff in the world. You are likely to find it quirky even when you run it in the environment it wants. Introducing more variables is just going to make problems more likely and your life more difficult.
Best thing is to simply try it. Doom 3 works with Vulkan. You may also see into commercial products based on Wine. Such exist. I heard a few apps work with OpenCL. Best thing would be to try. I assume it's not server cluster we are talking about.
should you?
Just build a new Z-80 SoftCard https://en.wikipedia.org/wiki/...
Domestic spying is now "Benign Information Gathering"
Is Windows better than Linux. I personally don't care and don't harbour a grudge if you use either. Windows offers easy installs and has a wider ( ? ) choice. Windows users just want their application or program to work on their computer instantly.. Linux (outside of Web searches, LibreOffice, and Emails, requires work and sometimes sweat if you are trying to make something work that doesn't. Why am I saying all this? (No! I don't want you to change your Operating System!) But.. If Windows decides to do something that makes you want to look elsewhere. Are you prepared to jump to a different ship if need be? Plus maybe to them, the threat of you jumping ship might keep them in line!
> Just run it in the OS it wants to run in on the iron.
Sometimes that certainly makes sense. Sometimes that would be really inconvenient and provide no benefit at all.
> Well any emulation/virtualization/whatever will slow things down. How much varies, but even in the best case scenario, there's a performance penalty.
That's not true. There is no measurable performance penalty for any application that's waiting on hardware such as the GPU or waiting on an external resource, typically over the network. As a stupid simple example, consider curl. Downloading a 100MB file over a 10Mbps connection will take 100 seconds, whether you're running on bare metal or three levels of virtualization deep.
If the CPU is sitting around waiting for the GPU to process the data, it doesn't matter if the CPU waits 10% "slower". A GPU-bound application will run just as fast whether the CPU is virtualized with PCI passthrough or not - it's not waiting on the CPU anyway.
Performance is all about the bottleneck. If the application is slowed because it's waiting on the CPU, it doesn't matter how fast the disk is. If the application spends it's time waiting on the network, it doesn't matter how fast the CPU is, etc. The only component that matters for performance is the component that's at ~100% utilization. (Ignoring for a moment trading disk and RAM via caching and paging).
You can absolutely run it if a VM is an accepted solution. The things you have to do:
1. Use a processor that supports virtualization
2. Have a second GPU for the host (in a server, usually the Melanox GPU).
3. Use a hypervisor that can pass through PCI ports and pass in the nVidia Cuda enabled GPU (like an nVidia Tesla M60 card)
4. Run your application in the VM
The real question is: why does the underlying system have to be Linux? Do you want to Hypervise it for HA? Without knowing the why's behind your request, it's hard to say whether you can or not. Is it absolutely possible? Yes. Is it the best approach? Maybe. Is it the most efficient approach for your high-spec'd application? Probably not.
There are cross platform libraries that work equally well on Linux and Windows.
You need to plan for it in advance.
Will running Windows Server really be that horrible for 1 server?
It may, it really depends. If you are running a system that must be compliant (PCI, NISPOM/JFAN, FedRAMP, etc..) it absolutely may be horrible. You not only need that 1 Windows server, but all of the supporting infrastructure to make it compliant. That 1 server is now 2 or 3, and you need manpower and time to go with it.
So if you can run Windows, I agree with GP. Where I work, the vendor would either port or we wouldn't run it.
-The wise argue that there are few absolutes, the fool argues that there are no probabilities.
My recent appempt to run the Windows software for High Fidelity in a Virtual Box on a linux machine with a good Nvidia card met with frustration when I chould not get the program to see the card from within the box.
My attempts to run the windows software for High Fidelity in a virtual box on a linux machine with a good Nvidia card were frustrated because the program could not use the video card.
I've been toying with the idea of moving my windows desktop to a VM inside a hypervisor. People have done passthrough on their gaming rigs but still take a 10-20% performance hit on the GPU.
Only the State obtains its revenue by coercion. - Murray Rothbard
I don't think a good hypervisor gives you the same drawback. You should get near-real performance on pass through devices now.
You are absolutely right that you can do this but you may need to patch the hypervisor if you use a standard, retail nVidia card. Nvidia's drivers for "gamer" GPUs check to see whether they are running in a virtual environment and then shutdown if they think they are. There is a patch for KVM which hides the virtual environment from the nVidia driver so that it will run.
Alternatively if you need double precision or don't mind spending 5-10 times the cost of a gamer card you can purchase a Tesla GPU since the nVidia driver for these works fine in a VM without patching.
The answer is yes and there's no need for a second GPU
Install Windows in VirtualBox or VMware, Configure the virtual machine for 3D and let it rip.
As for Wine, Wine is always win-some-lose-some. If the application will run under Wine, then it should have access to most of the Windows APIs of interest.
Finally, for CUDA, you would either need a second card or you would need an NVidia Grid solution which installs NVidia on the hypervisor and then again on the VM. It's insanely expensive.
If a VM is the accepted solution, the answer is still "No Linux cannot run a gpu computing application written for Windows."
A VM running Windows running the gpu application is still Windows running the gpu application.
That's what I meant, using the CUDA GPU for Open-CL or GPGPU type applications. In this case you need the second GPU so the system doesn't hold the GPU, freeing it up for the hypervisor.
Could you build a docker container based on Windows and run it in linux?
I've always said English was my second language. Had Romeo and Juliet been written in C, I might have understood it.
I'm not 100% sure, but I believe that it would run using docker. Docker I believe uses the same base engine that VirtualBox uses, although from my limited use of this great tool, it runs significantly faster with less drain on resources. I don't believe that it would have any problems with accessing the GPU - at least no more than any other virtualization system, so would be a viable option to consider.
Given nearly all the worlds fastest computers are Linux-based CUDA machines, you're full of it.
Grow some balls and post as yourself and I'll bother to answer why you're so wrong.