Managed Slurm for ML teams.
Share one GPU cluster across your team with managed Slurm from Jarvislabs. We set it up and run the scheduler. Compute nodes are full VMs connected through InfiniBand, with one shared filesystem on every node.
Now in beta
Scheduling
Queues and batch jobs, built in.
Slurm queues and runs your team's batch jobs, and tracks GPU use per team and per person.
Managed
A ready Slurm cluster, run by Jarvislabs.
Create a cluster from the console. Jarvislabs sets up Slurm, the network, the shared filesystem and the logins for your people, and runs the scheduler for you.
Ready to use
Slurm, containers, a login VM and a shared filesystem come set up. A cluster is usually ready in 10 to 15 minutes.
The scheduler is ours to run
Slurm and its job history run on a scheduler VM that Jarvislabs manages. You never log in to it.
Change it from the console
Add people, add or delete compute nodes, restart a compute node, and edit partitions and Slurm settings.
Your team
One cluster for the whole team.
Up to 50 people per cluster, each with a home folder on the shared filesystem.
People and roles
Add people with their SSH keys in the console. Admins get sudo on the login VM and every compute node. Users run jobs and work in their own home folder.
One shared folder
Code, data and checkpoints in /home/jl_fs are on the login VM and every compute node. The filesystem stays when you delete the cluster.
Partitions
Admins make partitions to keep compute nodes for a team or for long jobs.
Infrastructure
Real VMs, connected through InfiniBand.
Every node is a full VM. All compute nodes of a cluster are on one InfiniBand network.
Multi-node training
Run PyTorch with torchrun, srun or accelerate across compute nodes, and MPI with srun --mpi=pmix. A venv uses InfiniBand with no extra settings.
Containers built in
Pyxis and Enroot are set up. Pass a Docker Hub image to srun with --container-image.
GPUs today
Each compute node has 8 H200 GPUs and 8 InfiniBand cards. More GPU types are coming.
Specifications
The cluster, in one table.
What every cluster comes with.
Managed Slurm specifications
- Scheduler
- Details
- Slurm 26.05.5, on a scheduler VM that Jarvislabs manages
- Scheduling
- Details
- Queues, batch jobs and job arrays, with usage per team and per person
- Compute nodes
- Details
- Up to 128 VMs. Add or delete from the console
- Network
- Details
- 8 InfiniBand cards per compute node. All compute nodes on one InfiniBand network
- GPUs
- Details
- 8 H200 GPUs per compute node today. Up to 1,024 per cluster
- Cluster VMs
- Details
- A login VM, a scheduler VM and the compute nodes
- Containers
- Details
- Pyxis and Enroot, with Docker Hub images
- Shared files
- Details
- /home/jl_fs on the login VM and every compute node
- People
- Details
- Up to 50 per cluster, as admin (sudo) or user
| Setting | Details |
|---|---|
| Scheduler | Slurm 26.05.5, on a scheduler VM that Jarvislabs manages |
| Scheduling | Queues, batch jobs and job arrays, with usage per team and per person |
| Compute nodes | Up to 128 VMs. Add or delete from the console |
| Network | 8 InfiniBand cards per compute node. All compute nodes on one InfiniBand network |
| GPUs | 8 H200 GPUs per compute node today. Up to 1,024 per cluster |
| Cluster VMs | A login VM, a scheduler VM and the compute nodes |
| Containers | Pyxis and Enroot, with Docker Hub images |
| Shared files | /home/jl_fs on the login VM and every compute node |
| People | Up to 50 per cluster, as admin (sudo) or user |
Pricing
Pay by the hour for each VM.
You see the price of the whole cluster before you create it.
See the price first
The create page shows the price of your cluster per hour and per day.
Hourly, per VM
The login VM, the scheduler VM and each compute node bill by the hour.
Shared filesystem
The shared filesystem bills separately from the VMs of the cluster.
Before you start
Common questions.
Practical details for choosing and using this product.
Why Slurm for an ML team?
Slurm gives your team one queue for its GPUs. Jobs wait their turn and start when GPUs are free, and usage is tracked per team and per person. Nobody has to book GPUs by hand.
What do I get?
A Slurm 26.05.5 cluster: a login VM for your people, a scheduler VM that Jarvislabs manages, and compute nodes with up to 1,024 GPUs on one InfiniBand network, with a shared filesystem at /home/jl_fs on every node.
How long does a cluster take to be ready?
Usually 10 to 15 minutes after you create it.
How big can a cluster be?
Up to 1,024 GPUs, on up to 128 compute nodes. You can add or delete compute nodes from the console while the cluster runs.
Which GPUs can I use?
H200 today, with 8 GPUs per compute node. More GPU types are coming.
Does multi-node training use InfiniBand?
Yes. PyTorch training in a venv uses InfiniBand across nodes with no extra settings. In a container, the image needs the InfiniBand library. The docs show how to build one.
Can I run Docker images?
Yes. Pyxis and Enroot are set up, so srun --container-image=... works with Docker Hub images.
Who can log in?
The people you add in the console, with their SSH keys. They log in to the login VM with SSH. There are two roles: admin, with sudo, and user. Up to 50 people per cluster.
What happens to my files when I delete the cluster?
The shared filesystem at /home/jl_fs stays, with each person's home folder in it. It bills separately from the cluster.
Where does it run?
In the India-02 region.
Start your Slurm cluster.
Pick up to 1,024 GPUs. Your cluster is usually ready in 10 to 15 minutes.