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 SlurmMixin(object):
  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 _set_node_state(self, nodename, state, *args):
  20         cmd = ['scontrol', 'update', 'NodeName=' + nodename,
  21                'State=' + state]
  22         cmd.extend(args)
  23         subprocess.check_output(cmd)
  24
  25     def _get_slurm_state(self, nodename):
  26         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', nodename])
  27
  28
  29 class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
  30     def on_start(self):
  31         arv_node = self._arvados_node()
  32         if arv_node is None:
  33             self._nodename = None
  34             return super(ComputeNodeShutdownActor, self).on_start()
  35         else:
  36             self._set_logger()
  37             self._nodename = arv_node['hostname']
  38             self._logger.info("Draining SLURM node %s", self._nodename)
  39             self._later.issue_slurm_drain()
  40
  41     @RetryMixin._retry((subprocess.CalledProcessError,))
  42     def cancel_shutdown(self, reason):
  43         if self._nodename:
  44             if self._get_slurm_state(self._nodename) in self.SLURM_DRAIN_STATES:
  45                 # Resume from "drng" or "drain"
  46                 self._set_node_state(self._nodename, 'RESUME')
  47             else:
  48                 # Node is in a state such as 'idle' or 'alloc' so don't
  49                 # try to resume it because that will just raise an error.
  50                 pass
  51         return super(ComputeNodeShutdownActor, self).cancel_shutdown(reason)
  52
  53     @RetryMixin._retry((subprocess.CalledProcessError,))
  54     def issue_slurm_drain(self):
  55         if self.cancel_reason is not None:
  56             return
  57         if self._nodename:
  58             self._set_node_state(self._nodename, 'DRAIN', 'Reason=Node Manager shutdown')
  59             self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
  60             self._later.await_slurm_drain()
  61         else:
  62             self._later.shutdown_node()
  63
  64     @RetryMixin._retry((subprocess.CalledProcessError,))
  65     def await_slurm_drain(self):
  66         if self.cancel_reason is not None:
  67             return
  68         output = self._get_slurm_state(self._nodename)
  69         if output in ("drng\n", "alloc\n", "drng*\n", "alloc*\n"):
  70             self._timer.schedule(time.time() + 10,
  71                                  self._later.await_slurm_drain)
  72         elif output in ("idle\n"):
  73             # Not in "drng" so cancel self.
  74             self.cancel_shutdown("slurm state is %s" % output.strip())
  75         else:
  76             # any other state.
  77             self._later.shutdown_node()