Nsight: run CUDA code on remote gpu server

nsight, an eclipse IDE for CUDA programming

Steps for configuring the execution on remote gpu server:

---

Insallation:

1. nvidia cuda-toolkit can be installed from here. Not necessary to have GPU device on local.
2. set the system path for nvcc compiler.
3. run nvcc -V on you local machine to get the version.
4. run the nsight on you local, it will open the nsight ide in GUI.
5. ssh into remote the server and pull the docker image on remote server using: docker pull kayush206/ssh-docker
6. run the docker image on remote server using: nvidia-docker run -d -p <port>:22 kayush206/ssh-docker. replace the <port> with the available port of remote(i.e 54321).

Configuration:

1. Create a CUDA C/C++ Project:
   

2. Choose the project location on local as default and project type as Empty Project and click on Next.
   

3. Under the Basic settings, tick the Generate PTX code and Generate GPU code as 5.0 boxes and untick the 2.0 boxes (not supported with cuda-9.0) and click on Next.
   

4. Under Target Systems,click on manage and then click on add to add the remote connect details. replace the:

   • <ip_addr of gpu server> by gpu host address,
   • <port> by port to which container is mapped(i.e. 54321),
   • set User name to root,
   • set Label to gpu-container

   and click on Finish.
   

5. Under Target Systems, Add the project location and Toolkit details for remote(gpu-container) And selcet the x86_64 as CPU Architecture in Local System and gpu-container.

6. Copy all the *.cu and *.h files from the src to project.

7. Click on the Run Configuration from Run for configuring project in order to execute it on remote gpu server.

8. Right click on C/C++ Remote Application and choose New:

9. Click on the Remote tab and in Remote Connection choose gpu-container. And verify that Remote toolkit is exactly same as /usr/local/cuda/bin.
   Edit the Remote executable by clicking Run remote executable to /root/nsight-workspace/hello_project/Debug/hello_project after done with editing, click back to Upload local executable.
   

10. In order to synchronize your project files between local and remote, select the Set Active to gpu-container.
    And then, whenever you want to manually sync projecct files, click on Sync Active Now.

11. In order to build the project, choose Clean... and then Build All from Project tab.

12. click on the run button (green circle). if everything goes fine you will be able to see the output as details for you gpu devices.

References:

1. Executing CUDA C++ Application on a remote machine using Nsight Eclipse