Enable extra scheduler options through API config

Initial solution for #75.

cerise/back_end/job_runner.py

@@ -30,6 +30,8 @@ def __init__(self, job_store, config, api_files_path, api_install_script_path):
         """The name of the remote queue to submit jobs to."""
         self._mpi_slots_per_node = config.get_slots_per_node()
         """Number of MPI slots per node to request."""
+        self._scheduler_options = config.get_scheduler_options()
+        """Additional scheduler options to add."""
 
         self._logger.debug('Slots per node set to ' + str(self._mpi_slots_per_node))
 
@@ -107,6 +109,9 @@ def start_job(self, job_id):
             if self._queue_name:
                 jobdesc.queue_name = self._queue_name
 
+            if self._scheduler_options:
+                jobdesc.extra_scheduler_options = self._scheduler_options
+
             print("Starting job: " + str(jobdesc))
             job.remote_job_id = self._sched.submit(jobdesc)
             self._logger.debug('Job submitted')

cerise/back_end/test/test_job_runner.py

@@ -21,6 +21,9 @@ def get_queue_name(self):
     def get_slots_per_node(self):
         return 1
 
+    def get_scheduler_options(self):
+        return None
+
     def get_remote_cwl_runner(self):
         return '$CERISE_API_FILES/cerise/cwltiny.py'
 

cerise/config.py

@@ -231,6 +231,16 @@ def get_slots_per_node(self):
             return default
         return self._cr_config['jobs'].get('slots-per-node', default)
 
+    def get_scheduler_options(self):
+        """Returns the additional scheduler options to use.
+
+        Returns:
+            (str): The options as a single string.
+        """
+        if 'jobs' not in self._cr_config:
+            return None
+        return self._cr_config['jobs'].get('scheduler-options', None)
+
     def get_remote_refresh(self):
         """
         Returns the interval in between checks of the remote job \

docs/source/configuration.rst

@@ -131,6 +131,7 @@ format::
 
       queue-name: None      # cluster default
       slots-per-node: None  # cluster default
+      scheduler-options: None
       cwl-runner: $CERISE_API_FILES/cerise/cwltiny.py
 
     refresh: 10
@@ -172,9 +173,12 @@ Other valid values for ``scheduler`` are ``slurm``, ``torque`` and
 If jobs need to be sent to a particular queue, then you can pass the queue name
 using the corresponding option; if it is not specified, the default queue is
 used. If one or more of your steps start MPI jobs, then you may want to set the
-number of MPI slots per node via ``slots-per-node`` for better performance.
-Ideally, it would be possible to specify this in the CWL file for the step, but
-neither CWL nor Cerise currently support this.
+number of MPI slots per node via ``slots-per-node`` for better performance. If
+you need to specify additional scheduler options to e.g. select a GPU node, you
+can do so using e.g. ``scheduler-options: "-C TitanX --gres=gpu:1"``. Ideally,
+it would be possible to specify this in the CWL file for the step, but support
+for this in CWL is partial and in-development, and Cerise does not currently
+support this.
 
 Finally, ``cwl-runner`` specifies the remote path to the CWL runner. It defaults
 to ``$CERISE_API_FILES/cerise/cwltiny.py``, which is Cerise's included simple

integration_test/api/config.yml

@@ -15,3 +15,4 @@ compute-resource:
       username: xenon
       password: javagat
     scheduler: slurm
+    scheduler-options: --comment='testing'