CAIS++ GPU Server Manager

Managing scripts and information for the CAIS++ GPU servers

Server Connection Info

Dick: 68.181.99.174
- Two TitanX GPUs
- 1TB SSD
- i7 7700K CPU
- 32GB RAM
Bruce: 68.181.99.165
- Two TitanX GPUs
- 1TB SSD
- i7 7700K CPU
- 32GB RAM

Usage

If you don't know what you're doing with working on remote machines please see the "Getting Started" section first.

General

Please follow these rules.

Contact Andrew Szot if you need sudo for anything.
Running programs:
- If you are running a deep learning program always check that nobody else is using the GPU you want to use. Use the command nvidia-smi to check GPU usage.
- Only run your programs on one GPU This may require changing your code a little bit. Check this post out for how to use only certain GPUs keras-team/keras#6031.
- The bottleneck of both systems is the CPU. Don't run jobs that hog all of the CPU. Be careful about the number of cores you are using on some of the sklearn functions.
Anaconda:
- Every user shares the same Anaconda distribution. That means everyone has same environments. So if you're using an environment at the same time as someone else watch out for installs.
- All servers have a conda environment named tf do NOT install any packages to this environment. This environment is a base keras/tensorflow environment. I hope to keep it updated with everything a project should need.
- All servers also have a conda environment named pt for PyTorch.
- Don't use virtualenv or pip install anything without being in a conda environment.
Data
- The folders that you should be using are /home/yourusername (the starting directory), /hdd/datasets and /hdd/users/yourusername.
- /hdd/datasets is a public place to keep datasets for everyone to use. For instance if you want to use the CIFAR dataset first check that it doesn't already exist in /hdd/datasets and if doesn't download it there so other people can use it once you're done. If you're doing some sort of data preprocessing and need to save an altered version of a dataset my idea was that you would save that to /hdd/yourusername. This directory was intended to be a place for you to keep datasets specific to you.

I'm sure problems will come up with people all using the same computer so if you're having any difficulties just contact me.

Getting Started

These GPUs are avaliable for use on CAIS++ projects. The first thing you need is a user account. If you need a user account just message Andrew Szot on Slack with the username that you want along with your ssh public key. You will be able to log into all of the CAIS++ machines with this same account (you don't need to specify a password because of your ssh key).

Connecting to the Machine

Only read this section if you don't know how to work on a remote machine

Now that you have your username you are ready to connect to a machine. Go to the "servers" section of this README for the list of servers that you can connect to. Maybe check the "servers" slack channel for server avaliability. Now you have an IP that you want to connect to.

There are a couple of different ways of actually working and running scripts on a remote machine.

The easiest way is to get a SFTP client. This will allow you to transfer files on your compute to the remote computer over SFTP (file transfer over SSH). For Mac I recommend Cyberduck. For Windows I recommend Filezilla. You should be able to connect and see the remote machine file system as well as your own file system on both of these programs. I am not very familiar with Filezilla but I know that on Cyberduck you can edit files on the remote computer's filesystem from a local text editor like sublime or whatever you use. Or you can edit the files on your computer and then drag and drop them onto the server filesystem.

I would recommend not downloading big datasets to your personal computer and then uploading them to the remote machine using SFTP. Use wget.

Now that your files are on the server you have to log onto the server to actually run whatever programs you made and see the output. You have to log into the server using SSH. If you're on Mac or Linux just use the SSH command (ssh yourusername@serverip). You will then be prompted for your password. Of course if you want you can add your public key on the remote machine so you don't have to worry about typing your password every time. On Windows use PuTTY. Now that you're connected remotely to the server you can use the command line like you normally would.

An issue you will run into is everytime you close your laptop the SSH session will close and your program will stop running. We want a way to keep the program running even when you are logged out. For that use tmux session.