X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/b5d3c559df7250ab7106adc2ded7f38dc666cf9c..0eb72b526bf8bbb011551ecf019f604e17a534f1:/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py

diff --git a/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py b/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
index 71e73f17f6..fa56578cff 100644
--- a/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
+++ b/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
@@ -1,4 +1,7 @@
 #!/usr/bin/env python
+# Copyright (C) The Arvados Authors. All rights reserved.
+#
+# SPDX-License-Identifier: AGPL-3.0
 
 from __future__ import absolute_import, print_function
 
@@ -6,46 +9,88 @@ import subprocess
 import time
 
 from . import \
-    ComputeNodeSetupActor, ComputeNodeUpdateActor, ComputeNodeMonitorActor
+    ComputeNodeSetupActor, ComputeNodeMonitorActor
 from . import ComputeNodeShutdownActor as ShutdownActorBase
+from . import ComputeNodeUpdateActor as UpdateActorBase
+from .. import RetryMixin
 
-class ComputeNodeShutdownActor(ShutdownActorBase):
-    SLURM_END_STATES = frozenset(['down\n', 'down*\n', 'drain\n', 'fail\n'])
+class SlurmMixin(object):
+    SLURM_END_STATES = frozenset(['down\n', 'down*\n',
+                                  'drain\n', 'drain*\n',
+                                  'fail\n', 'fail*\n'])
+    SLURM_DRAIN_STATES = frozenset(['drain\n', 'drng\n'])
 
+    def _set_node_state(self, nodename, state, *args):
+        cmd = ['scontrol', 'update', 'NodeName=' + nodename,
+               'State=' + state]
+        cmd.extend(args)
+        subprocess.check_output(cmd)
+
+    def _get_slurm_state(self, nodename):
+        return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', nodename])
+
+
+class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
     def on_start(self):
         arv_node = self._arvados_node()
         if arv_node is None:
+            self._nodename = None
             return super(ComputeNodeShutdownActor, self).on_start()
         else:
+            self._set_logger()
             self._nodename = arv_node['hostname']
             self._logger.info("Draining SLURM node %s", self._nodename)
             self._later.issue_slurm_drain()
 
-    def _set_node_state(self, state, *args):
-        cmd = ['scontrol', 'update', 'NodeName=' + self._nodename,
-               'State=' + state]
-        cmd.extend(args)
-        subprocess.check_output(cmd)
-
-    @ShutdownActorBase._retry((subprocess.CalledProcessError,))
-    def cancel_shutdown(self):
-        self._set_node_state('RESUME')
-        return super(ComputeNodeShutdownActor, self).cancel_shutdown()
+    @RetryMixin._retry((subprocess.CalledProcessError, OSError))
+    def cancel_shutdown(self, reason, try_resume=True):
+        if self._nodename:
+            if try_resume and self._get_slurm_state(self._nodename) in self.SLURM_DRAIN_STATES:
+                # Resume from "drng" or "drain"
+                self._set_node_state(self._nodename, 'RESUME')
+            else:
+                # Node is in a state such as 'idle' or 'alloc' so don't
+                # try to resume it because that will just raise an error.
+                pass
+        return super(ComputeNodeShutdownActor, self).cancel_shutdown(reason)
 
-    @ShutdownActorBase._stop_if_window_closed
-    @ShutdownActorBase._retry((subprocess.CalledProcessError,))
+    @RetryMixin._retry((subprocess.CalledProcessError, OSError))
     def issue_slurm_drain(self):
-        self._set_node_state('DRAIN', 'Reason=Node Manager shutdown')
-        self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
-        self._later.await_slurm_drain()
+        if self.cancel_reason is not None:
+            return
+        if self._nodename:
+            self._set_node_state(self._nodename, 'DRAIN', 'Reason=Node Manager shutdown')
+            self._logger.info("Waiting for SLURM node %s to drain", self._nodename)
+            self._later.await_slurm_drain()
+        else:
+            self._later.shutdown_node()
 
-    @ShutdownActorBase._stop_if_window_closed
-    @ShutdownActorBase._retry((subprocess.CalledProcessError,))
+    @RetryMixin._retry((subprocess.CalledProcessError, OSError))
     def await_slurm_drain(self):
-        output = subprocess.check_output(
-            ['sinfo', '--noheader', '-o', '%t', '-n', self._nodename])
-        if output in self.SLURM_END_STATES:
-            self._later.shutdown_node()
-        else:
+        if self.cancel_reason is not None:
+            return
+        output = self._get_slurm_state(self._nodename)
+        if output in ("drng\n", "alloc\n", "drng*\n", "alloc*\n"):
             self._timer.schedule(time.time() + 10,
                                  self._later.await_slurm_drain)
+        elif output in ("idle\n"):
+            # Not in "drng" but idle, don't shut down
+            self.cancel_shutdown("slurm state is %s" % output.strip(), try_resume=False)
+        else:
+            # any other state.
+            self._later.shutdown_node()
+
+    def _destroy_node(self):
+        if self._nodename:
+            self._set_node_state(self._nodename, 'DOWN', 'Reason=Node Manager shutdown')
+        super(ComputeNodeShutdownActor, self)._destroy_node()
+
+
+class ComputeNodeUpdateActor(UpdateActorBase):
+    def sync_node(self, cloud_node, arvados_node):
+        if arvados_node.get("hostname"):
+            try:
+                subprocess.check_output(['scontrol', 'update', 'NodeName=' + arvados_node["hostname"], 'Weight=%i' % int(cloud_node.size.price * 1000)])
+            except:
+                self._logger.error("Unable to set slurm node weight.", exc_info=True)
+        return super(ComputeNodeUpdateActor, self).sync_node(cloud_node, arvados_node)