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 subprocess32 as 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 _update_slurm_size_attrs(self, nodename, size):
  32         self._update_slurm_node(nodename, [
  33             'Weight=%i' % int(size.price * 1000),
  34             'Features=instancetype=' + size.id,
  35         ])
  36
  37     def _get_slurm_state(self, nodename):
  38         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', nodename])
  39
  40
  41 class ComputeNodeSetupActor(SlurmMixin, SetupActorBase):
  42     def create_cloud_node(self):
  43         hostname = self.arvados_node.get("hostname")
  44         if hostname:
  45             self._update_slurm_size_attrs(hostname, self.cloud_size)
  46         return super(ComputeNodeSetupActor, self).create_cloud_node()
  47
  48
  49 class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
  50     def on_start(self):
  51         arv_node = self._arvados_node()
  52         if arv_node is None:
  53             self._nodename = None
  54             return super(ComputeNodeShutdownActor, self).on_start()
  55         else:
  56             self._set_logger()
  57             self._nodename = arv_node['hostname']
  58             self._logger.info("Draining SLURM node %s", self._nodename)
  59             self._later.issue_slurm_drain()
  60
  61     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  62     def cancel_shutdown(self, reason, try_resume=True):
  63         if self._nodename:
  64             if try_resume and self._get_slurm_state(self._nodename) in self.SLURM_DRAIN_STATES:
  65                 # Resume from "drng" or "drain"
  66                 self._update_slurm_node(self._nodename, ['State=RESUME'])
  67             else:
  68                 # Node is in a state such as 'idle' or 'alloc' so don't
  69                 # try to resume it because that will just raise an error.
  70                 pass
  71         return super(ComputeNodeShutdownActor, self).cancel_shutdown(reason)
  72
  73     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  74     def issue_slurm_drain(self):
  75         if self.cancel_reason is not None:
  76             return
  77         if self._nodename:
  78             self._update_slurm_node(self._nodename, [
  79                 'State=DRAIN', 'Reason=Node Manager shutdown'])
  80             self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
  81             self._later.await_slurm_drain()
  82         else:
  83             self._later.shutdown_node()
  84
  85     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  86     def await_slurm_drain(self):
  87         if self.cancel_reason is not None:
  88             return
  89         output = self._get_slurm_state(self._nodename)
  90         if output in ("drng\n", "alloc\n", "drng*\n", "alloc*\n"):
  91             self._timer.schedule(time.time() + 10,
  92                                  self._later.await_slurm_drain)
  93         elif output in ("idle\n",):
  94             # Not in "drng" but idle, don't shut down
  95             self.cancel_shutdown("slurm state is %s" % output.strip(), try_resume=False)
  96         else:
  97             # any other state.
  98             self._later.shutdown_node()
  99
 100     def _destroy_node(self):
 101         if self._nodename:
 102             self._update_slurm_node(self._nodename, [
 103                 'State=DOWN', 'Reason=Node Manager shutdown'])
 104         super(ComputeNodeShutdownActor, self)._destroy_node()
 105
 106
 107 class ComputeNodeUpdateActor(SlurmMixin, UpdateActorBase):
 108     def sync_node(self, cloud_node, arvados_node):
 109         """Keep SLURM's node properties up to date."""
 110         hostname = arvados_node.get("hostname")
 111         features = arvados_node.get("slurm_node_features", "").split(",")
 112         sizefeature = "instancetype=" + cloud_node.size.id
 113         if hostname and sizefeature not in features:
 114             # This probably means SLURM has restarted and lost our
 115             # dynamically configured node weights and features.
 116             self._update_slurm_size_attrs(hostname, cloud_node.size)
 117         return super(ComputeNodeUpdateActor, self).sync_node(
 118             cloud_node, arvados_node)