Instructions for using Vision Lab Machines/ Multi-GPU Servers

Compiled by: Soumen Basu


  1. We recommend installing Anaconda for installing packages. Use the required cudatoolkit in your conda environments. You can create multiple environments. DO NOT try to install cuda on your own. It may cause problems for other users using the same system. In almost all cases you can use a new conda environment with the required cudatoolkit version. In case you absolutely need an update of the cuda version then please contact the Lab Admin

    NOTE: You are not given sudo access for the same reason. In case you need any package installation then contact the Lab Admin.


  2. If you are using the multi-GPU servers, then please use the server GPUs responsibly. Don’t hog all the GPUs unnecessarily and use a SINGLE GPU unless you got approval from your advisor for priority access for a critical task. 

    NOTE: You may use multiple GPUs if they are free. However, you must ensure that the other users are not starved. As a rule of thumb, inform all users if you are going to use more than 2 GPUs for more than 4 hours (with your Estimated Time to Completion). Similarly, announce if you got/ need priority access along with ETC for a critical task.


  3. To connect to the internet from your terminal use the proxy scripts. More details can be found here (Courtesy: Anupam Sobti). This is an internal link. Using screen/ tmux sessions is a good idea for working remotely.

  4. Since the disk space is limited, you will get 35 GB of home space by default. However, you may use the scratch folder designated for you for keeping the larger datasets/ files/ folders. A link to your designated scratch folder should be present in your home folder.

    NOTE: Create all large datasets/ files inside scratch, and not at home. Otherwise, you will end up using your home quota even if you move it to scratch later.


  5. The quota system allows you for an additional 15GB for 7 days whenever you exceed the 35 GB quota. That means, in case you have exceeded the 35GB, you get a grace period of 7 days to clean up your home space to bring it under 35 GB. However, if you have not cleaned up space within the grace period, your disk space will behave as it is completely out of space and you will not be able to create any new files/folders.

  6. If you are working in the vision0X/ cse0X servers, then please take regular backup of your scratch data in local machines/ HPC/ Onedrive or any other persistent storage of your choice. You can request up to 2TB free home space in HPC through your advisor. Onedrive provides 1 TB free space for IITD users.

    7. NOTE: vision0X/ cse0X servers are for compute purpose and not for storage. Your files and folders in the scratch are liable to be deleted if the disk space starts to get full. We will try to delete the older files on a best effort basis (no-guarantees).


  7. Multi-GPU server (vision0X/ cse0X) server users must use the designated MS Teams channel for inter-user communication and raising issues to the admin.

  8. DO NOT download large datasets directly in the Multi-GPU (vision0X/ cse0X) servers. Download the dataset in your local machine/ HPC/ Lab's RAID Server (negroni) first and then transfer to the vision0X/ cse0X servers. You may consider using the research proxy for bulk download. You can request it from CSC website and you advisor would approve it. Set the environment variables as, http_proxy="http://xen03.iitd.ac.in:3128" and https_proxy="http://xen03.iitd.ac.in:3128" in your bash terminal. The credentials of the RAID server is available in our MS Teams group.