Accelerator code and MS remote desktop

Hi

I recently found out that accelerator programs will not run on Windows servers using remote desktop - a bit of a showstopper. Using VNC is not really an option either …

Does anyone know whether Microsofts upcoming remoteFX GPU additions to Remote desktop will solve this issue?

Regards,

Casper

Hi Casper,

I definitely feel your pain with regard to remote desktop, VNC and GPU computing. One way I have worked around it is by running an SSH server in Cygwin. If you have PGI Workstation installed you are already half-way there; You will just need to download the Cygwin installer and get the packages listed below. You can then extract them to the directory where the PGI install placed Cygwin and follow any number of online tutorials for setting it up (googling ‘sshd cygwin’ is a good place to start). Once you are logged on to a Windows system via SSH you will have full access to the GPU.

This is the first I have heard of remoteFX and after reading about it I’m not overly optimistic it will solve this problem. As soon as I can get a copy of SP1 for Server 2008 R2 I will try it out and try to post some information on the user forum.

Packages needed to add sshd to PGI Workstation:

binutils
crypt
csih
cygrunsrv
editrights
libwrap
minires
openssh
openssl

We are true Windows folks around here, so we run PVF and not the workstation edition … Installing cygwin is not really an option anyway, since multiple other people in academia use our binaries in combination with remote desktop …

Are you aware of any workarounds? Having some workaround program running at the console would solve the problem, since servers are rebooted very rarely …

I attended Michael Wolfes short course on GPU programming in Hamburg last weekend, which was nothing short of awesome and really inspired me. I already figured out how to speed up our code by a factor of ~5, however, the remote desktop issue remains a showstopper.

I’m afraid the only options are VNC and SSH. I am holding out hope that remoteFX will solve the problem for remote desktop but it looks like it is still a ways off.

If you develop via remote desktop and save your binaries to the remote system, it is fairly easy to log in via SSH when it comes time to run them ;)

There are several stand-alone SSH servers for windows that don’t require a Cygwin install, i.e. CopSSH

Another possibility is to use PVF locally, and save builds to a network share. I imagine that you could then add some sort of custom build rule to log in to the development system via SSH and execute the binary. Just thinking off the top of my head here; I talked to a PVF engineer and she seemed to think it was possible.

Server 2008 R2 sp1 beta is out:

http://www.microsoft.com/windowsserver2008/en/us/sp1.aspx

Perhaps you could give it a try…?

Sorry for the late reply.

Unfortunately RemoteFX does not allow remote GPU access at the level required for CUDA or PGI Accelerator/CUDA Fortran. However there is certainly better access to physical hardware (I was even able to install the NVIDIA drivers over remote desktop).

You may be able to get more information from Microsoft’s SP1 Beta forum:

http://social.technet.microsoft.com/Forums/en/w7itproSP/threads

Thanks … I was really hoping for this to work though …

I can live with using VNC or having to work locally while programming and debugging, but for executing the final binary remote desktop compatibility is a neccesity. I guess this would be the case for many other people as well …

Would it be possible to make a windows service that can be called with the path to an accelerator executable which is then launched in local user space, preferably returning an exit code when done? I am just trying to think outside the box here …

Casper

It just came to my attention that Nvidia released special Tesla Compute Cluster(TCC) drivers, which solves the remote desktop issue:

http://forums.nvidia.com/index.php?showtopic=159208

From what I can tell, support for TSS is improved in CUDA toolkit 3.2:

http://developer.nvidia.com/object/cuda_3_2_toolkit_rc.html
http://www.anandtech.com/show/3924/nvidia-announces-parallel-nsight-15-cuda-toolkit-32

Does this mean remote desktop will be supported in a future PVF release using the 3.2 toolkit?

It just came to my attention that Nvidia released special Tesla Compute Cluster(TCC) drivers, which solves the remote desktop issue:

Thanks Casper. I’ll let Doug know once he’s back from vacation.

Does this mean remote desktop will be supported in a future PVF release using the 3.2 toolkit?

PVF can be run via a remote desktop now. What can’t be run via remote desktop is GPU code due to a limitation in the NVIDIA driver. If the new TCC drivers do solve this problem, then I don’t see any reason why you shouldn’t be able to run GPU programs from current versions of PVF.

As for new CUDA tool-kits, we generally add them to the first PGI release after they are officially released (i.e. non-beta).

Thanks again for the information.
Mat