Karolina supercomputer

• Karolina has three different storages:

  • Home storage: small capacity
  • Project storage: large capacity, slow, persistent
  • Scratch storage: large capacity, fast, regularly cleared

• All storages are shared across all nodes of the supercomputer.

• It is recommended to store simulation data in the Project storage and submit jobs from there. Data on the Project storage are not deleted until the project finishes, and this storage has a large capacity (unlike Home storage).

• Prefer to not submit jobs from the Scratch storage, as its content is regularly deleted and you can lose your data.

• With default options, when submitting from the Project storage, qq automatically creates a directory on Scratch storage, executes your job there, and transfers the data back. qq takes care of these copying operations.

• If you want to know more about configuring the storage qq uses for your job, read this section of the manual.

• Submit jobs with the --account option providing your project ID. You can find your project ID by running it4ifree (left-most column, in the format OPEN-12-34).

• When submitting CPU-only jobs (queues starting with qcpu), you always need to allocate a full compute node. Each CPU node has 128 CPU cores. If you do not specify the number of CPUs, qq will use the correct value automatically.

• For most Gromacs simulations, it is recommended to simulate multiple systems as part of a single node-wide job. You can use the qq_loop_re or the qq_flex_re run scripts for that.

• When submitting CPU+GPU jobs (queues starting with qgpu), you can allocate as little as 1/8 of a compute node, which corresponds to 1 GPU and 16 CPU cores.

• Karolina's compute nodes have fast interconnections, making it easy to efficiently run jobs across multiple nodes.

Click here for detailed external documentation of the Karolina supercomputer.