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