tccp - Tufts Cluster Command Prompt

why this exists

Running code on an HPC cluster shouldn't feel different from running it on your laptop. Open your editor, change a file, run it on a GPU. But the usual workflow is: write a SLURM batch script, scp your code over, submit, check the queue, scp results back. Or use a web UI with a clunky browser notebook. tccp replaces all of that with a single CLI that makes the cluster feel local.

HPC clusters have real constraints. Your home directory is 30GB — that fills up fast once you start using datasets and installing packages. Scratch space is big but gets wiped. Figuring out what goes where, cleaning up after yourself, and not blowing your quota is a whole second job on top of the actual work.

tccp takes care of all of that for you! Just pretend you're local with JUMBO compute!

the experience

You work in your own editor, on your own machine. When you're ready to run, you type tccp start. tccp allocates a node, pulls your container, syncs your code, and gives you a persistent shell on the compute node with GPU access. You can detach, go get coffee, come back and reattach. If your WiFi dies, the shell keeps running. Press Ctrl+S to sync code changes and reattach instantly.

It feels like running locally, except the work is happening on cluster hardware.

compared to the alternatives

vs. writing SLURM scripts by hand

The traditional way: write an sbatch script, scp your code, submit it, check squeue, scp the results back, repeat. Every change means re-uploading files. You manage your own containers, environments, and storage. tccp does all of this for you in a single command.

vs. the On Demand web UI

The web UI gives you a Jupyter notebook in a browser or a remote desktop. It works, but you're stuck in their editor and their environment. You can't use your local tools, your dotfiles, your keybindings. tccp lets you stay in your own setup and just sends the work to the cluster.

what you get

quick start

# install the binary (see install page)

# save your credentials once
$ tccp setup

# set up your project
$ cd my-project
# create tccp.yaml with your container

# start a session and get a shell
$ tccp start
$ tccp shell
tccp> python train.py

v — built for the Tufts HPC cluster — source on GitHub — llms.txt