services/nodemanager/arvnodeman/computenode/dispatch/slurm.py

   1 #!/usr/bin/env python
   2
   3 from __future__ import absolute_import, print_function
   4
   5 import subprocess
   6 import time
   7
   8 from . import \
   9     ComputeNodeSetupActor, ComputeNodeUpdateActor, ComputeNodeMonitorActor
  10 from . import ComputeNodeShutdownActor as ShutdownActorBase
  11
  12 class ComputeNodeShutdownActor(ShutdownActorBase):
  13     SLURM_END_STATES = frozenset(['down\n', 'down*\n',
  14                                   'drain\n', 'drain*\n',
  15                                   'fail\n', 'fail*\n'])
  16
  17     def on_start(self):
  18         arv_node = self._arvados_node()
  19         if arv_node is None:
  20             self._nodename = None
  21             return super(ComputeNodeShutdownActor, self).on_start()
  22         else:
  23             self._nodename = arv_node['hostname']
  24             self._logger.info("Draining SLURM node %s", self._nodename)
  25             self._later.issue_slurm_drain()
  26
  27     def _set_node_state(self, state, *args):
  28         cmd = ['scontrol', 'update', 'NodeName=' + self._nodename,
  29                'State=' + state]
  30         cmd.extend(args)
  31         subprocess.check_output(cmd)
  32
  33     @ShutdownActorBase._retry((subprocess.CalledProcessError,))
  34     def cancel_shutdown(self):
  35         if self._nodename:
  36             self._set_node_state('RESUME')
  37         return super(ComputeNodeShutdownActor, self).cancel_shutdown()
  38
  39     @ShutdownActorBase._stop_if_window_closed
  40     @ShutdownActorBase._retry((subprocess.CalledProcessError,))
  41     def issue_slurm_drain(self):
  42         self._set_node_state('DRAIN', 'Reason=Node Manager shutdown')
  43         self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
  44         self._later.await_slurm_drain()
  45
  46     @ShutdownActorBase._stop_if_window_closed
  47     @ShutdownActorBase._retry((subprocess.CalledProcessError,))
  48     def await_slurm_drain(self):
  49         output = subprocess.check_output(
  50             ['sinfo', '--noheader', '-o', '%t', '-n', self._nodename])
  51         if output in self.SLURM_END_STATES:
  52             self._later.shutdown_node()
  53         else:
  54             self._timer.schedule(time.time() + 10,
  55                                  self._later.await_slurm_drain)