University of Hamburg - to research, to teach, to educate and form, to homepage
RRZ
RRZ

Photo: Thomas Orgis

RRZ tools

rrz-batch-login

rrz-batch-login can be employed to requests a login session on a compute node under control of the batch system. (This implies that batch job limits apply, and that one might have to wait until resources occupied by other batch jobs become available.)

Contents

Typical usage

Command line Action
rrz-batch-login request a login session in the std partition
rrz-batch-login --gpus=1 request a login session in the gpu partition
rrz-batch-login --jobid=jobID request a session sharing resources with job jobID

Use cases

rrz-batch-login provides a convenient way to interactive use of compute nodes. The advantage over ssh is that ssh cannot allocate resources. One can only ssh into a node. But this will only be allowed if an own batch job is running on that node. The ssh session will be put into the cgroup of a running job. If more that one own batch job is running the assignment of the cgroup will be arbitrary. In any case, an ssh session would share resources with a batch job.

There are two use cases:

  • Exclusive interactive access to compute resources. The normal place for interactive work are the front-end nodes. rrz-batch-login shall only be employed if the front-end nodes cannot accomplish a task. For example, this can happen when testing parallel CPU programs or when a GPU is needed.

  • Monitoring of batch jobs. If --jobid is given the interactive session will be in the cgroup of the corresponding batch job. In particular, it is possible to check how much memory is used for Temporary/scratch files. Note, however, that the df command gives wrong numbers, one has to resort to du or ls -l.

Full description

The full command line reads:

rrz-batch-login [--dryrun] [Option] [srun-options]

The --help option can be used to get a short description of the command:

rrz-batch-login --help

Meaning of the other parameters:

  • --dryrun will print the srun command generated but does not execute it.
  • Option can be
    • the number of GPUs to be requested: --gpus=n
    • a batch job-ID: --jobid=jobID
  • srun-options are parameters passed to srun (check with --dryrun).