Computational work on Vista is scheduled through the Slurm Workload Manager. You use Slurm commands to submit, monitor, and control your jobs, even interactive ones. Jobs submitted to the scheduler are queued, then run on the compute nodes when resources become available. Vista's compute nodes are inaccessible to you until they have been allocated to your job by Slurm.

Running on the login nodes violates the TACC Good Conduct Guide mentioned earlier, and it may even result in your account being suspended. Therefore:

• Never launch a resource-heavy executable directly from the command line of a login node
• Note this policy applies to building your codes, too: make -j 4 is fine, but make -j 64 isn't

In this topic, we focus on the process of submitting jobs to run on the compute nodes. In the subsequent Managing Jobs topic, we look at how to monitor and adjust your jobs after they have been submitted.

Job Accounting

In Vista’s accounting system, 1 Service Unit (SU) equals three hours of use on a CPU compute node or one hour of use on a GPU compute node.

For example, a job that runs on 4 GG nodes for 3 hours will cost 4 SUs: \[4 \text{ nodes} \times 3 \text{ hours} \times 0.33 = 4 \text{ SUs}\]

Note that the accounting system will only charge you for the resources you are allocated while your job actually runs, not for the time length you requested. All running jobs will be charged a minimum of 15 minutes regardless of actual job runtime. Still, your job will generally spend less time waiting in the queue if you set a time limit that is fairly close to what you truly need for the job. The scheduler will find it easier to come up with a slot for the 2 hours you really need instead of the 24 hours you might have requested arbitrarily in your job script.

To display a summary of your TACC project balances and disk quotas at any time, execute this on a login node: