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, ComputeNodeMonitorActor
  10 from . import ComputeNodeShutdownActor as ShutdownActorBase
  11 from . import ComputeNodeUpdateActor as UpdateActorBase
  12 from .. import RetryMixin
  13
  14 class SlurmMixin(object):
  15     SLURM_END_STATES = frozenset(['down\n', 'down*\n',
  16                                   'drain\n', 'drain*\n',
  17                                   'fail\n', 'fail*\n'])
  18     SLURM_DRAIN_STATES = frozenset(['drain\n', 'drng\n'])
  19
  20     def _set_node_state(self, nodename, state, *args):
  21         cmd = ['scontrol', 'update', 'NodeName=' + nodename,
  22                'State=' + state]
  23         cmd.extend(args)
  24         subprocess.check_output(cmd)
  25
  26     def _get_slurm_state(self, nodename):
  27         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', nodename])
  28
  29
  30 class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
  31     def on_start(self):
  32         arv_node = self._arvados_node()
  33         if arv_node is None:
  34             self._nodename = None
  35             return super(ComputeNodeShutdownActor, self).on_start()
  36         else:
  37             self._set_logger()
  38             self._nodename = arv_node['hostname']
  39             self._logger.info("Draining SLURM node %s", self._nodename)
  40             self._later.issue_slurm_drain()
  41
  42     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  43     def cancel_shutdown(self, reason, try_resume=True):
  44         if self._nodename:
  45             if try_resume and self._get_slurm_state(self._nodename) in self.SLURM_DRAIN_STATES:
  46                 # Resume from "drng" or "drain"
  47                 self._set_node_state(self._nodename, 'RESUME')
  48             else:
  49                 # Node is in a state such as 'idle' or 'alloc' so don't
  50                 # try to resume it because that will just raise an error.
  51                 pass
  52         return super(ComputeNodeShutdownActor, self).cancel_shutdown(reason)
  53
  54     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  55     def issue_slurm_drain(self):
  56         if self.cancel_reason is not None:
  57             return
  58         if self._nodename:
  59             self._set_node_state(self._nodename, 'DRAIN', 'Reason=Node Manager shutdown')
  60             self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
  61             self._later.await_slurm_drain()
  62         else:
  63             self._later.shutdown_node()
  64
  65     @RetryMixin._retry((subprocess.CalledProcessError, OSError))
  66     def await_slurm_drain(self):
  67         if self.cancel_reason is not None:
  68             return
  69         output = self._get_slurm_state(self._nodename)
  70         if output in ("drng\n", "alloc\n", "drng*\n", "alloc*\n"):
  71             self._timer.schedule(time.time() + 10,
  72                                  self._later.await_slurm_drain)
  73         elif output in ("idle\n"):
  74             # Not in "drng" but idle, don't shut down
  75             self.cancel_shutdown("slurm state is %s" % output.strip(), try_resume=False)
  76         else:
  77             # any other state.
  78             self._later.shutdown_node()
  79
  80     def _destroy_node(self):
  81         if self._nodename:
  82             self._set_node_state(self._nodename, 'DOWN', 'Reason=Node Manager shutdown')
  83         super(ComputeNodeShutdownActor, self)._destroy_node()
  84
  85
  86 class ComputeNodeUpdateActor(UpdateActorBase):
  87     def sync_node(self, cloud_node, arvados_node):
  88         if arvados_node.get("hostname"):
  89             try:
  90                 subprocess.check_output(['scontrol', 'update', 'NodeName=' + arvados_node["hostname"], 'Weight=%i' % int(cloud_node.size.price * 1000)])
  91             except:
  92                 self._logger.error("Unable to set slurm node weight.", exc_info=True)
  93         return super(ComputeNodeUpdateActor, self).sync_node(cloud_node, arvados_node)