X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/15b49783aa0ac76508986e772b98ffa9d187c57f..8a41cc44ee196c9347785baa476a370abe77c75c:/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 cbeabd1a8a..5b7785afd9 100644
--- a/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
+++ b/services/nodemanager/arvnodeman/computenode/dispatch/slurm.py
@@ -1,13 +1,17 @@
 #!/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
 
-import subprocess
+import subprocess32 as subprocess
 import time
 
-from . import \
-    ComputeNodeSetupActor, ComputeNodeUpdateActor, ComputeNodeMonitorActor
+from . import ComputeNodeMonitorActor
+from . import ComputeNodeSetupActor as SetupActorBase
 from . import ComputeNodeShutdownActor as ShutdownActorBase
+from . import ComputeNodeUpdateActor as UpdateActorBase
 from .. import RetryMixin
 
 class SlurmMixin(object):
@@ -16,16 +20,32 @@ class SlurmMixin(object):
                                   '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 _update_slurm_node(self, nodename, updates):
+        cmd = ['scontrol', 'update', 'NodeName=' + nodename] + updates
+        try:
+            subprocess.check_output(cmd)
+        except:
+            self._logger.error(
+                "SLURM update %r failed", cmd, exc_info=True)
+
+    def _update_slurm_size_attrs(self, nodename, size):
+        self._update_slurm_node(nodename, [
+            'Weight=%i' % int(size.price * 1000),
+            'Features=instancetype=' + size.id,
+        ])
 
     def _get_slurm_state(self, nodename):
         return subprocess.check_output(['sinfo', '--noheader', '-o', '%t', '-n', nodename])
 
 
+class ComputeNodeSetupActor(SlurmMixin, SetupActorBase):
+    def create_cloud_node(self):
+        hostname = self.arvados_node.get("hostname")
+        if hostname:
+            self._update_slurm_size_attrs(hostname, self.cloud_size)
+        return super(ComputeNodeSetupActor, self).create_cloud_node()
+
+
 class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
     def on_start(self):
         arv_node = self._arvados_node()
@@ -38,30 +58,31 @@ class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
             self._logger.info("Draining SLURM node %s", self._nodename)
             self._later.issue_slurm_drain()
 
-    @RetryMixin._retry((subprocess.CalledProcessError,))
+    @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')
+                self._update_slurm_node(self._nodename, ['State=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)
 
-    @RetryMixin._retry((subprocess.CalledProcessError,))
+    @RetryMixin._retry((subprocess.CalledProcessError, OSError))
     def issue_slurm_drain(self):
         if self.cancel_reason is not None:
             return
         if self._nodename:
-            self._set_node_state(self._nodename, 'DRAIN', 'Reason=Node Manager shutdown')
+            self._update_slurm_node(self._nodename, [
+                'State=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()
 
-    @RetryMixin._retry((subprocess.CalledProcessError,))
+    @RetryMixin._retry((subprocess.CalledProcessError, OSError))
     def await_slurm_drain(self):
         if self.cancel_reason is not None:
             return
@@ -69,9 +90,29 @@ class ComputeNodeShutdownActor(SlurmMixin, ShutdownActorBase):
         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"):
+        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._update_slurm_node(self._nodename, [
+                'State=DOWN', 'Reason=Node Manager shutdown'])
+        super(ComputeNodeShutdownActor, self)._destroy_node()
+
+
+class ComputeNodeUpdateActor(SlurmMixin, UpdateActorBase):
+    def sync_node(self, cloud_node, arvados_node):
+        """Keep SLURM's node properties up to date."""
+        hostname = arvados_node.get("hostname")
+        features = arvados_node.get("slurm_node_features", "").split(",")
+        sizefeature = "instancetype=" + cloud_node.size.id
+        if hostname and sizefeature not in features:
+            # This probably means SLURM has restarted and lost our
+            # dynamically configured node weights and features.
+            self._update_slurm_size_attrs(hostname, cloud_node.size)
+        return super(ComputeNodeUpdateActor, self).sync_node(
+            cloud_node, arvados_node)