nikola-sur / gpufun

Interactive GPU prototyping on Sockeye

Setup SSH

1. Setup a password manager to avoid having to type your CWL password every time
2. Make sure Sockeye connection sharing is setup. See relevant section of Sockeye doc. The key info is that you should have the following in your laptop's ~/.ssh/config file (~ mean home directory):

Host s
    HostName sockeye.arc.ubc.ca
    User mycwl
    ControlMaster auto
    ControlPath ~/.ssh/ssh_mux_%h_%p_%r

3. Open the UBC VPN and a terminal window and login with ssh s, do the 2-factor auth.
4. Put that terminal aside and don't touch it. As long as it is open, other terminals can be opened without the 2 step auth (required for VS Code remote to work).

Setup remote VS Code

1. Open VS Code and click on the lower right icon >< to start a remote VS Code running on Sockeye.
2. For the hostname, just type s
3. Once you are in, use the terminal in VS Code, open it with apple/control with "`"

Basic Sockeye setup

1. Create the directory /scratch/st-alexbou-1/[your username]
2. Download julia and install in your home dir, e.g. in /home/[your username]/bin/julia-1.xx.x
3. Edit ~/.bash_profile to add:

export PATH=$PATH:/home/[your username]/bin/julia-1.10.2/bin

export JULIA_PKG_DEVDIR=/scratch/st-alexbou-1/[your username]/
export JULIA_DEPOT_PATH=/scratch/st-alexbou-1/[your username]/depot

4. Test julia with julia then you will need this single package globally ] add Revise (others are better to install in a project specific manner) then exit()
5. Edit ~/.bashrc to add

alias j='julia --banner=no --load /home/[your username]/julia-start.jl --project=@. ' 
alias interact='salloc --time=3:00:00 --mem=16G --nodes=1 --ntasks=2 --account=st-alexbou-1'
alias ginteract='salloc --account=st-alexbou-1-gpu --partition=interactive_gpu --time=1:0:0 -N 1 -n 2 --mem=8G --gpus=1'

6. Create ~/julia-start.jl

println("Active project: $(Base.active_project())")
println("Loading Revise...")
using Revise

Setup this repo

1. Clone this repo, i.e. cd /scratch/st-alexbou-1/[your username], git clone https://github.com/alexandrebouchard/gpufun.git, cd in it.
2. Open two terminals in VS code:
   • One will be in the head node. Use for anything that needs network access (package install in particular). Open julia with j
   • One will be a GPU node that does not have internet access. In that terminal, type ginteract to start the GPU node, then j.
3. In the head node terminal, setup the Julia dependencies with ] instantiate

Test your setup

In the GPU node's julia: include("test.jl")