Skip to content

Compute nodes and queues

Request a compute node, choose a queue and priority, and give the node back when you are done. A compute node is a computer you use for a set time: HPC Hub asks Slurm for one in the size you choose and shows the time left on the cluster card.

Get a compute node

Turn on Get a compute node when I connect to request a node every time you connect. When that setting is off, request a node yourself:

  1. Open the window with Get one on the cluster card, or with Compute node in the ⋮ menu.
  2. Set the size, cores, memory, time, queue, and priority.
  3. Select Get node.

The node is a Slurm job, so it also shows on Jobs. It may wait in the queue before it starts.

What you can set

Size is a starting point: Light CPU (2 cores, 8 GB), Heavy CPU (8 cores, 32 GB), or GPU. You can then change Cores, Memory, and Time, up to what the queue allows.

Queue and priority is covered in Queues and priorities.

Cost is the SUs the request costs, on clusters that bill for them.

Queues and priorities

Queue and priority has three choices. HPC Hub reads them from Slurm after you connect.

Queue
Which nodes and time limits the job uses. Queues of GPU nodes are listed with the GPU size. Light CPU and Heavy CPU leave those queues out. Queues that only other groups can use are hidden.
Priority
How the job is billed, and how soon it starts.
Group to bill
The account the hours go to. Leave it empty to use your default group.

HPC Hub keeps the choices consistent:

  • A queue that takes only its own priority (a debug queue, for example) switches the priority to that one and tells you.
  • A queue of GPU nodes adds a GPU to the request and switches the size to GPU.
  • A combination Slurm would never start cannot be sent.

Learn about queues, at the top of the queue list, opens the Limits table.

Whole-node queues

On some queues you always get whole nodes. The queue's line says "You always get whole nodes, even if you ask for fewer cores", and the cost is counted per node. Pick a queue like that when you can use the whole machine.

Some queues also say "Other jobs can stop yours". Save your work often on those queues.

Time left

The card shows your node and the time left, for example: On node n0123, job 123456, 3 h 20 min left.

About 15 minutes before the end, HPC Hub sends a notification. When the time is up, Slurm ends the job, and everything on that node stops, including its terminals.

"started elsewhere" means the node came from another job you already had. That is the Reuse a node you already have setting.

Give a node back

Release node ends the job early. Anything running on the node stops, and you stay connected.

  1. Open the ⋮ menu on the card.
  2. Select Release node.
  3. Confirm when HPC Hub asks.

Disconnect releases the node too. Quitting the app keeps the node. Open Settings and select Disconnect all and quit to disconnect and quit.

The maintenance warning

Slurm starts a job only when that job can finish before booked maintenance begins. If the time you ask for runs into maintenance, the window says so before you send the request:

Warning

Maintenance starts Tue, Nov 10 at 7:50 AM EST. A 12 h job can't finish before then, so it would wait until maintenance is over. Shortened, it can start now.

Do one of the following:

  • Select Shorten to, with a time that ends before maintenance. The job can start now.
  • Select Keep 12 h and wait (the button shows your time). The job waits until maintenance is over.

The time stays as you set it until you shorten it.