Glossary
These pages use one term for each thing. Slurm's own terms and the names on console pages stay as they are.
Parts of a cluster
| Term | Meaning |
|---|---|
| Cluster | A set of VMs that Jarvislabs builds for you: one scheduler VM, one login VM and your compute nodes. |
| Scheduler VM | The VM that runs Slurm's own services. Jarvislabs manages it. You cannot log in to it. |
| Login VM | The VM that you log in to. You submit jobs from it. |
| Compute node | A VM with 8 H200 GPUs and 8 InfiniBand cards. Jobs run on it. Slurm commands call it a node. |
| Shared filesystem | The storage at /home/jl_fs. The login VM and all compute nodes see the same files. |
| Home folder | Your own folder on the shared filesystem: /home/jl_fs/<login-name>. In commands, ~ means this folder. For example, for user1, ~/results is /home/jl_fs/user1/results. |
| Console | The Jarvislabs website, where you create and manage clusters. |
People
| Term | Meaning |
|---|---|
| Person | Someone on the cluster's People tab. A person has one role: admin or user. |
| Login name | The name a person uses to log in. It also names the home folder. You cannot change it later. |
| Admin | A person who can use sudo, and change partitions, compute node states, accounts and limits from the login VM. |
| User | A person who can submit jobs and control their own jobs. |
Slurm terms
| Term | Meaning |
|---|---|
| Slurm | The software that queues your programs and gives them GPUs. |
| Job | A program that you ask Slurm to run. |
| Job script | A file with #SBATCH lines and the commands of a job. You give it to sbatch. |
| Partition | A named group of compute nodes with its own limits. You start with one, gpu, with all compute nodes and no time limit. |
| Drain | A state of a compute node: it takes no new jobs, and its jobs finish. |
| Resume | The command that puts a drained compute node back in service. |
| Requeue | Slurm puts a job back in the queue. It runs again from the start. |
| Prolog, epilog | A script that Slurm runs before a job (prolog) or after a job (epilog). |
| Account | A Slurm group of people. You use it to report on a team and limit it. |
| QOS | Quality of service: a named set of limits that you give to people. |
| Fair share | A Slurm rule that gives a higher priority to people who used less of the cluster. |
| Rank | The number of one process among all processes of a job, from 0. |
| gres | Slurm's word for GPUs in a job. |
Software
| Term | Meaning |
|---|---|
| InfiniBand | The fast network between compute nodes. |
| Ethernet | The normal, slower network. |
| NCCL | NVIDIA's library that moves data between GPUs. PyTorch uses it. |
| DDP | PyTorch's way to train across several GPUs. |
| MPI | A way to run one program as many processes across compute nodes. |
| PMIx | The way MPI processes find each other under Slurm. |
| Container | Software packed with all that it needs, so it runs the same anywhere. |
| Pyxis, Enroot | The tools that run containers under Slurm. srun options such as --container-image use them. |
| Venv | A Python virtual environment: a folder with its own Python packages. |
| uv | A tool that installs Python and Python packages. |
| Checkpoint | A file where your program saves its progress, so it can continue later. |
| SSH rules | The security group rules that control who can reach the login VM's public IP address. |