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

   1 #!/usr/bin/env python
   2 # Copyright (C) The Arvados Authors. All rights reserved.
   3 #
   4 # SPDX-License-Identifier: AGPL-3.0
   5
   6 from __future__ import absolute_import, print_function
   7
   8 import subprocess
   9 import time
  10
  11 from . import \
  12     ComputeNodeSetupActor, ComputeNodeMonitorActor
  13 from . import ComputeNodeShutdownActor as ShutdownActorBase
  14 from . import ComputeNodeUpdateActor as UpdateActorBase
  15 from .. import RetryMixin
  16
  17 class SlurmMixin(object):
  18     SLURM_END_STATES = frozenset(['down\n', 'down*\n',
  19                                   'drain\n', 'drain*\n',
  20                                   'fail\n', 'fail*\n'])
  21     SLURM_DRAIN_STATES = frozenset(['drain\n', 'drng\n'])
  22
  23     def _set_node_state(self, nodename, state, *args):
  24         cmd = ['scontrol', 'update', 'NodeName=' + nodename,
  25                'State=' + state]
  26         cmd.extend(args)
  27         subprocess.check_output(cmd)
  28
  29     def _get_slurm_state(self, nodename):
  30         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', nodename])
  31
  32
  33 class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
  34     def on_start(self):
  35         arv_node = self._arvados_node()
  36         if arv_node is None:
  37             self._nodename = None
  38             return super(ComputeNodeShutdownActor, self).on_start()
  39         else:
  40             self._set_logger()
  41             self._nodename = arv_node['hostname']
  42             self._logger.info("Draining SLURM node %s", self._nodename)
  43             self._later.issue_slurm_drain()
  44
  45     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  46     def cancel_shutdown(self, reason, try_resume=True):
  47         if self._nodename:
  48             if try_resume and self._get_slurm_state(self._nodename) in self.SLURM_DRAIN_STATES:
  49                 # Resume from "drng" or "drain"
  50                 self._set_node_state(self._nodename, '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(reason)
  56
  57     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  58     def issue_slurm_drain(self):
  59         if self.cancel_reason is not None:
  60             return
  61         if self._nodename:
  62             self._set_node_state(self._nodename, '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         else:
  66             self._later.shutdown_node()
  67
  68     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  69     def await_slurm_drain(self):
  70         if self.cancel_reason is not None:
  71             return
  72         output = self._get_slurm_state(self._nodename)
  73         if output in ("drng\n", "alloc\n", "drng*\n", "alloc*\n"):
  74             self._timer.schedule(time.time() + 10,
  75                                  self._later.await_slurm_drain)
  76         elif output in ("idle\n"):
  77             # Not in "drng" but idle, don't shut down
  78             self.cancel_shutdown("slurm state is %s" % output.strip(), try_resume=False)
  79         else:
  80             # any other state.
  81             self._later.shutdown_node()
  82
  83     def _destroy_node(self):
  84         if self._nodename:
  85             self._set_node_state(self._nodename, 'DOWN', 'Reason=Node Manager shutdown')
  86         super(ComputeNodeShutdownActor, self)._destroy_node()
  87
  88
  89 class ComputeNodeUpdateActor(UpdateActorBase):
  90     def sync_node(self, cloud_node, arvados_node):
  91         if arvados_node.get("hostname"):
  92             try:
  93                 subprocess.check_output(['scontrol', 'update', 'NodeName=' + arvados_node["hostname"], 'Weight=%i' % int(cloud_node.size.price * 1000)])
  94             except:
  95                 self._logger.error("Unable to set slurm node weight.", exc_info=True)
  96         return super(ComputeNodeUpdateActor, self).sync_node(cloud_node, arvados_node)