Dear Shashikant,

I really doubt it works. The CUDA program must have access to the graphics driver, and from the description of PsExec it won't have. There may be other ways, if you have a Tesla card.

I have had a similar problem, and the best solution which worked great for me was the free TightVNC.

Hope it helps.

Dear Farshid,

Thank you for your answer, TightVNC looks like a remote desktop control software with GUI, I don't want to use remote machine interactively, I want to use remote machine GPU card to execute my CUDA executable file which resides in local machine,  transferring executable over network. is there any method you come across to resolve this issue. Kindly help me if possible. 

Dear Shashikant Ilager,

You can use OpenMPI to distribute processing into the network together, OpenMP to distribute over CPU cores, and assign a CPU core thread to a graphics CUDA device even remotely. The solution is OpenMPI (network)+OpenMP(inside a single machine) +CUDA (inside graphics device).

Hoping it helps.

Abel

Dear Abel Gomes,

Thank you for your suggestions. I would like to explore these options. Can we transfer an executable file in MPI and make it execute on a particular machine. Since MPI always schedules our tasks in all machine together.

Dear Shashikant Ilager;

Regarding your last question addressing MIP "kinda".

The idea is to have a MPI binary that via a black-box of encapsulation and communication (MPI), which is programming language independent (mostly for C and Fortran but supported by other languages), allows you to assign processing to specific local CPU cores, 1 or more, or to remote CPUs. Thus, you do not send the binary you (and I mean MPI here) sends the data that is required to process something remotely.

For a local machine MPI can do the same as OpenMP in terms of processing distribution among cores. However, MPI is more complex (requires much more code) to use than OpenMP.

Thus, usually MPI for communication among cluster/lan work distribution, OpenMP to assign work to several cores per remote machine and finally you use OpenCL (AMD or nVidia) or CUDA (nVidia) if you want to use the graphics card in each remote compute.

Here is a detailed example solely for MPI.

http://mpitutorial.com/tutorials/running-an-mpi-cluster-within-a-lan/

Dear Shashikant,

As I already mentioned, when it comes to remote execution of CUDA programs, you have very few choices. To run a CUDA code, the CUDA runtime needs to have access to graphics drivers, which are loaded only if you somehow grant your code to access them. The easiest way is the TightVNC solution, which enables remote copying of the files and also running them. Having said this, if you have a Tesla card, you can run the drivers in a special mode that removes this limitation, and you can use your favorite tool to remotely execute the code. Take a look at the attache link; read all comments too.

http://stackoverflow.com/questions/12304082/can-i-run-a-cuda-program-on-a-computer-if-i-remote-into-it-using-windows-rdp

Dear Shashikant,

The building purpose of CUDA is not to perform in distributed machines &/ connecting distributed machines for task distribution. GPU based system works best in instruction level parallelism. Data parallelism is best on distributed / network machines.  Thus, my advice to use Ninf-G for data distribution between distributed machines (GRID structure) then MPI for node task / data distribution and CUDA for task distribution inside each node. You also need to remember the communication overhead, which should not be over-passed the execution efficiency of the task parallelism.

Thanks.