Remote execution using Cygwin ssh

I have two Win7 PC’s where I develop on my host machine and run on a target machine with the CUDA board in it. I need to have it this way so that I can debug using Parallel Nsight. When I’m not debugging I would like to be able to just run from a command prompt. I have Cygwin sshd running on my taget machine and can ssh into it, but when I run my CUDA application I always get an “unspecified driver error”. Is there some way I can remotely run my applications on my target machine?

Thanks,
Peter

You can install linux system on the target machine. I installed ubuntu on my desktop and it’s working fine

You can install linux system on the target machine. I installed ubuntu on my desktop and it’s working fine

Thanks for the suggestion, but as far as I know there is no Nsight monitor for Linux. Before Nsight was available I had CentOS installed and used GDB for debugging, but I’d hoped that this new debug environment would be easier to use. It is easier, just has some quirks.

Peter

Thanks for the suggestion, but as far as I know there is no Nsight monitor for Linux. Before Nsight was available I had CentOS installed and used GDB for debugging, but I’d hoped that this new debug environment would be easier to use. It is easier, just has some quirks.

Peter

using TCC should solve this, I assume sshd is running as a service and as a result is affected by session 0 isolation

using TCC should solve this, I assume sshd is running as a service and as a result is affected by session 0 isolation

Yes, sshd is running as a service. What is TCC?

Peter

Yes, sshd is running as a service. What is TCC?

Peter

TCC = Tesla CLuster Computing.

Tesla devices have TCC drivers.

Those drivers allow you to use CUDA via RDesktop Session.

Unfortunately I wasn’t able to run GTX 470 in TCC mode. I made it to install Tesla drivers but it is not enough - simple inf modification.

I managed to patch sys file to return Tesla C2050 devid instead GTX 470 devid. Managed to patch nvidia-ism to pass -dm 1 for GTX 470.

However CUDA is still not working via RDP. I’m pretty much sure that this is software imposed limitation.

This is really sad because I have to use VNC or some other remote tool instead of RDP; or spend thousands of $$$$ for Tesla.

The last resort is to patch devid in graphic card BIOS - I plan to do that in next few days; I managed to locate devid bytes in BIOS.

My personal opinion is that nVidia won’t loose anything by enabling TCC for GTX nv100 series - more over nV can only benefit even more by enabling TCC for nv100.

regards

M.

TCC = Tesla CLuster Computing.

Tesla devices have TCC drivers.

Those drivers allow you to use CUDA via RDesktop Session.

Unfortunately I wasn’t able to run GTX 470 in TCC mode. I made it to install Tesla drivers but it is not enough - simple inf modification.

I managed to patch sys file to return Tesla C2050 devid instead GTX 470 devid. Managed to patch nvidia-ism to pass -dm 1 for GTX 470.

However CUDA is still not working via RDP. I’m pretty much sure that this is software imposed limitation.

This is really sad because I have to use VNC or some other remote tool instead of RDP; or spend thousands of $$$$ for Tesla.

The last resort is to patch devid in graphic card BIOS - I plan to do that in next few days; I managed to locate devid bytes in BIOS.

My personal opinion is that nVidia won’t loose anything by enabling TCC for GTX nv100 series - more over nV can only benefit even more by enabling TCC for nv100.

regards

M.

Vowie! That sounds like a cool thing to do…

But why are people averse to VNC?

Vowie! That sounds like a cool thing to do…

But why are people averse to VNC?

well, i don’t know about cygwin, but i am now using vncviewer.exe to remotely access and run my GPU equipped WINDOW 7 machine from WINDOW XP machine…

well, i don’t know about cygwin, but i am now using vncviewer.exe to remotely access and run my GPU equipped WINDOW 7 machine from WINDOW XP machine…

Finally determined that this was the error I was having:

Dell M4500

Win 7 64-bit

Quadro FX 1800M

Cygwin 1.7.9-1

Some of the world drudgingly use Windoze, especially those of us who are menu-adverse Luddites who prefer command-line compilation and scripting (and who frequently work in places where we are trying to get the customer up from stone knives and bear skins and menus are for order-out).

I recently got this hot new box, spec’d by IT and containing required IT-ish applications but would be handy for working on-the-road, and started the process of getting CUDA installed. Very disappointed: I’ve been using PuTTY to access the local Windoze box like it was any of the dozen Linux boxes that I use every day, including the embedded, rugged, deployable Linux box that is running a very interesting CUDA application. However, after much experimentation I determined that there was this permissions issue (“session 0” mentioned above) which was not going away.

Of course, that’s after installing and de-installing CUDA 4.0 x64 and 3.2 x64, currently working with 4.0 x32.

Basically no sshd access to CUDA on a Windoze box.

Had I known the right question to ask, I may have been here days ago. Thank you for verifying my findings.

PS: Linux CUDA has no such issue, although Linux ATI CAL actually does with default X Window Server configurations.

Research would seem to indicate that TCC is not appropriate for a Quadro FX 1800M either: did I get that right?

The Cygwin local bash shell, which uses the STUPID WINDOWS CONSOLE, so cuts and pastes are DIFFERENT and STUPID and CLUMSY AND ERROR PRONE, and command-line compilations get chewed up.

However, I have found that I can bring up a local xterm using the Cygwin X Window Server run from a bash shell in the STUPID WINDOWS CONSOLE, and THAT allows more UNIX-ish cut-and-paste.

I hope that helps somebody in the same soup.

Now I get to re-try CUDA 4.0 x64…

Thank God I will never spec a Windoze box for a customer…

AND NOW FOR SOMETHING COMPLETELY DIFFERENT!

Beware of what Cygwin bash shows you in Windows 7.

When building a CUDA example for the first time, you may run into the following message:

Custom build rules file ‘C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\VCProjectDefaults\NvCudaRuntimeApi.rules’ was not found or failed to load.

Apparently this was included in the SDK pre-3.2, but after 3.2 it was included in the CUDA Toolkit; one has to install it manually.

One should install it using Windoze tools, like cut-and-paste from Explorer, rather than copying it with bash under Cygwin.

Why? Because Windows 7 HIDES S**T.

Apparently, if the Administrator created a directory, then only files placed in that directory by Administrator are visible to Session 0 programs. You may be able to see it with Cygwin bash, but that doesn’t mean it’s there, really, to things like VS2008.

I performed the above necessary task copying from a Cygwin bash shell. Files looks like they were there, but the only diagnostic indicated that they weren’t.

Copied with Explorer, and all was well.

Hopefully, this posting will save 8 somebodies at least an hour of frustration to make up for my… error?

Experience is what you get when you didn’t get what you wanted…