SBATCH switch cheat sheet for Agave
The below switches can be used either with interactive or in an SBATCH script.
Switches can be used in combination with each other to optimize the resources assigned to a job
Command | Effect | Usage example | Notes |
|---|---|---|---|
-N <X> | Request X number of nodes for job to be spread across Default = 1 Default lets Slurm choose. Slurm will always try to make this 1 if possible combined with other options |
#SBATCH -N 2 | The system will strictly enforce this even if your job could be run on a smaller number of nodes. For non-MPI jobs this may make your job run slower e.g. specifying -N 2 -n 2 will result in getting one core on each of two compute nodes even though all nodes have multiple cores |
-n <X> | Create a job with If requested under the number of cores/system in a partition it will default -N to 1 If requested over the number of cores/system in the specified partition Slurm will find the fastest way to assign all requested cores |
#SBATCH -n 28 | Usually only MPI jobs require tasks |
-c <X> | Request X cores/task Default = 1 Similar to -n, however is multiplied by the number of specified tasks |
#SBATCH -c 4 |
|
-t days-hours or -t minutes or -t DD-HH:MM:SS | Requested amount of time for a job The default is 1 day (1-0) with a maximum of 7 days. Job will end after the specified amount of time if jobs in queue need the resources. |
| Requesting less time can reduce the amount of time a job waits before starting at the risk of your job ending before it is completed |
-p <partition> | Specify which partition to use. Default = (Depends) |
#SBATCH -p publicgpu | In most cases Slurm can determine this itself. |
-q <qos> | Specify the QOS to use (e.g. a queue) Default = “normal” |
| The QOS must be specified with a partition when the QOS is not the default, normal. |
--gres | Specify a “Generic Resource” to request with the job. Commonly used to request GPU’s Default = None See our GPU documentation for details |
#SBATCH -p gpu | Generally only used for GPUs, and since GPUs are only available with the |
-o <filename> | Specifies a filename to capture all standard output to. By itself, also captures standard error messages. This is equivalent to ‘ mycommand > filename.foo ' Default = None |
#SBATCH -o %j.out | Not recommended for interactive jobs. %j in example captures the job number for unique filename |
-e <filename> | Specifies filename to capture all standard error messages within. This is equivalent to ' mycommand 2>&1 > filename.foo ' Default = None |
#SBATCH -e %j.err | Not recommended for interactive jobs. %j in example captures the job number for unique filename |
--mail-type=<$VAR> | Specify when you want to receive an e-mail update about a job even such as starting, finishing, failure etc Default = None | interactive = N/A #SBATCH --mail-type=ALL | Not recommended for interactive jobs useful for status updates on SBATCH jobs |
--mail-user=%u@asu.edu | Email address to send notifications to. %u automatically expands to username | interactive = N/A #SBATCH --mail-user=%u@asu.edu | Not recommended for interactive jobs useful for status updates on SBATCH jobs |