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 ComputeNodeMonitorActor
  12 from . import ComputeNodeSetupActor as SetupActorBase
  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 _update_slurm_node(self, nodename, updates):
  24         cmd = ['scontrol', 'update', 'NodeName=' + nodename] + updates
  25         try:
  26             subprocess.check_output(cmd)
  27         except:
  28             self._logger.error(
  29                 "SLURM update %r failed", cmd, exc_info=True)
  30
  31     def _get_slurm_state(self, nodename):
  32         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', nodename])
  33
  34
  35 class ComputeNodeSetupActor(SlurmMixin, SetupActorBase):
  36     def create_cloud_node(self):
  37         hostname = self.arvados_node.get("hostname")
  38         if hostname:
  39             self._update_slurm_node(self.arvados_node['hostname'], [
  40                 'Weight=%i' % int(self.cloud_size.price * 1000),
  41                 'Features=instancetype='+self.cloud_size.name,
  42             ])
  43         return super(ComputeNodeSetupActor, self).create_cloud_node()
  44
  45
  46 class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
  47     def on_start(self):
  48         arv_node = self._arvados_node()
  49         if arv_node is None:
  50             self._nodename = None
  51             return super(ComputeNodeShutdownActor, self).on_start()
  52         else:
  53             self._set_logger()
  54             self._nodename = arv_node['hostname']
  55             self._logger.info("Draining SLURM node %s", self._nodename)
  56             self._later.issue_slurm_drain()
  57
  58     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  59     def cancel_shutdown(self, reason, try_resume=True):
  60         if self._nodename:
  61             if try_resume and self._get_slurm_state(self._nodename) in self.SLURM_DRAIN_STATES:
  62                 # Resume from "drng" or "drain"
  63                 self._update_slurm_node(self._nodename, ['State=RESUME'])
  64             else:
  65                 # Node is in a state such as 'idle' or 'alloc' so don't
  66                 # try to resume it because that will just raise an error.
  67                 pass
  68         return super(ComputeNodeShutdownActor, self).cancel_shutdown(reason)
  69
  70     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  71     def issue_slurm_drain(self):
  72         if self.cancel_reason is not None:
  73             return
  74         if self._nodename:
  75             self._update_slurm_node(self._nodename, [
  76                 'State=DRAIN', 'Reason=Node Manager shutdown'])
  77             self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
  78             self._later.await_slurm_drain()
  79         else:
  80             self._later.shutdown_node()
  81
  82     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  83     def await_slurm_drain(self):
  84         if self.cancel_reason is not None:
  85             return
  86         output = self._get_slurm_state(self._nodename)
  87         if output in ("drng\n", "alloc\n", "drng*\n", "alloc*\n"):
  88             self._timer.schedule(time.time() + 10,
  89                                  self._later.await_slurm_drain)
  90         elif output in ("idle\n",):
  91             # Not in "drng" but idle, don't shut down
  92             self.cancel_shutdown("slurm state is %s" % output.strip(), try_resume=False)
  93         else:
  94             # any other state.
  95             self._later.shutdown_node()
  96
  97     def _destroy_node(self):
  98         if self._nodename:
  99             self._update_slurm_node(self._nodename, [
 100                 'State=DOWN', 'Reason=Node Manager shutdown'])
 101         super(ComputeNodeShutdownActor, self)._destroy_node()
 102
 103
 104 class ComputeNodeUpdateActor(SlurmMixin, UpdateActorBase):
 105     def sync_node(self, cloud_node, arvados_node):
 106         hostname = arvados_node.get("hostname")
 107         if hostname:
 108             self._update_slurm_node(hostname, [
 109                 'Weight=%i' % int(cloud_node.size.price * 1000),
 110                 'Features=instancetype=' + cloud_node.size.name,
 111             ])
 112         return super(ComputeNodeUpdateActor, self).sync_node(
 113             cloud_node, arvados_node)