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     SLURM_DRAIN_STATES = frozenset(['drain\n', 'drng\n'])
  17
  18     def on_start(self):
  19         arv_node = self._arvados_node()
  20         if arv_node is None:
  21             self._nodename = None
  22             return super(ComputeNodeShutdownActor, self).on_start()
  23         else:
  24             self._nodename = arv_node['hostname']
  25             self._logger.info("Draining SLURM node %s", self._nodename)
  26             self._later.issue_slurm_drain()
  27
  28     def _set_node_state(self, state, *args):
  29         cmd = ['scontrol', 'update', 'NodeName=' + self._nodename,
  30                'State=' + state]
  31         cmd.extend(args)
  32         subprocess.check_output(cmd)
  33
  34     def _get_slurm_state(self):
  35         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', self._nodename])
  36
  37     # The following methods retry on OSError.  This is intended to mitigate bug
  38     # #6321 where fork() of node manager raises "OSError: [Errno 12] Cannot
  39     # allocate memory" resulting in the untimely death of the shutdown actor
  40     # and tends to result in node manager getting into a wedged state where it
  41     # won't allocate new nodes or shut down gracefully.  The underlying causes
  42     # of the excessive memory usage that result in the "Cannot allocate memory"
  43     # error are still being investigated.
  44
  45     @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
  46     def cancel_shutdown(self):
  47         if self._nodename:
  48             if self._get_slurm_state() in self.SLURM_DRAIN_STATES:
  49                 # Resume from "drng" or "drain"
  50                 self._set_node_state('RESUME')
  51             else:
  52                 # Node is in a state such as 'idle' or 'alloc' so don't
  53                 # try to resume it because that will just raise an error.
  54                 pass
  55         return super(ComputeNodeShutdownActor, self).cancel_shutdown()
  56
  57     @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
  58     @ShutdownActorBase._stop_if_window_closed
  59     def issue_slurm_drain(self):
  60         self._set_node_state('DRAIN', 'Reason=Node Manager shutdown')
  61         self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
  62         self._later.await_slurm_drain()
  63
  64     @ShutdownActorBase._retry((subprocess.CalledProcessError, OSError))
  65     @ShutdownActorBase._stop_if_window_closed
  66     def await_slurm_drain(self):
  67         output = self._get_slurm_state()
  68         if output in self.SLURM_END_STATES:
  69             self._later.shutdown_node()
  70         else:
  71             self._timer.schedule(time.time() + 10,
  72                                  self._later.await_slurm_drain)