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', 'drain\n', 'fail\n'])
  14
  15     def on_start(self):
  16         arv_node = self._arvados_node()
  17         if arv_node is None:
  18             self._nodename = None
  19             return super(ComputeNodeShutdownActor, self).on_start()
  20         else:
  21             self._nodename = arv_node['hostname']
  22             self._logger.info("Draining SLURM node %s", self._nodename)
  23             self._later.issue_slurm_drain()
  24
  25     def _set_node_state(self, state, *args):
  26         cmd = ['scontrol', 'update', 'NodeName=' + self._nodename,
  27                'State=' + state]
  28         cmd.extend(args)
  29         subprocess.check_output(cmd)
  30
  31     @ShutdownActorBase._retry((subprocess.CalledProcessError,))
  32     def cancel_shutdown(self):
  33         if self._nodename:
  34             self._set_node_state('RESUME')
  35         return super(ComputeNodeShutdownActor, self).cancel_shutdown()
  36
  37     @ShutdownActorBase._stop_if_window_closed
  38     @ShutdownActorBase._retry((subprocess.CalledProcessError,))
  39     def issue_slurm_drain(self):
  40         self._set_node_state('DRAIN', 'Reason=Node Manager shutdown')
  41         self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
  42         self._later.await_slurm_drain()
  43
  44     @ShutdownActorBase._stop_if_window_closed
  45     @ShutdownActorBase._retry((subprocess.CalledProcessError,))
  46     def await_slurm_drain(self):
  47         output = subprocess.check_output(
  48             ['sinfo', '--noheader', '-o', '%t', '-n', self._nodename])
  49         if output in self.SLURM_END_STATES:
  50             self._later.shutdown_node()
  51         else:
  52             self._timer.schedule(time.time() + 10,
  53                                  self._later.await_slurm_drain)