def process_pbsnode(self, pbsnode, instance_ids_to_shutdown, nodearray_definitions):
'''
If the pbsnode is offline, will handle evaluating whether the node can be shutdown. See instance_ids_to_shutdown, which
is an OUT parameter here.
Otherwise convert the pbsnode into a cyclecloud.machine.Machine instance.
'''
states = set(pbsnode["state"].split(","))
resources = pbsnode["resources_available"]
# host has incorrect case
hostname = resources["vnode"]
instance_id = resources.get("instance_id", autoscale_util.uuid("instanceid"))
def try_shutdown_pbsnode():
if not instance_id:
pbscc.error("instance_id was not defined for host %s, can not shut it down" % hostname)
elif "down" in states:
# don't immediately remove down nodes, give them time to recover from network failure.
remove_down_nodes = float(self.cc_config.get("pbspro.remove_down_nodes", 300))
since_down = self.clock.time() - pbsnode["last_state_change_time"]
if since_down > remove_down_nodes:
pbscc.error("Removing down node %s after %.0f seconds", hostname, since_down)
instance_ids_to_shutdown[instance_id] = hostname
return True
else:
omega = remove_down_nodes - since_down
pbscc.warn("Not removing down node %s for another %.0f seconds", hostname, omega)
else:
instance_ids_to_shutdown[instance_id] = hostname
return True
return False
if "offline" in states:
if not pbsnode.get("jobs", []):
pbscc.fine("%s is offline and has no jobs, may be able to shut down" % hostname)
if try_shutdown_pbsnode():
return
else:
pbscc.fine("Host %s is offline but still running jobs" % hostname)
# if the node is just in the down state, try to shut it down.
if set(["down"]) == states and try_shutdown_pbsnode():
return
# just ignore complex down nodes (down,job-busy etc) until PBS decides to change the state.
if "down" in states:
return
# convert relevant resources from bytes to floating point (GB)
for key in resources:
value = resources[key]
if isinstance(value, basestring) and value.lower() in ["true", "false"]:
value = value.lower() == "true"
elif isinstance(value, list):
# TODO will need to support this eventually
continue
else:
try:
resources[key] = pbscc.parse_gb_size(key, resources[key])
except InvalidSizeExpressionError:
pass
resources["hostname"] = hostname
nodearray_name = resources.get("nodearray") or resources.get("slot_type")
group_id = resources.get("group_id")
if resources.get("machinetype") and nodearray_name:
machinetype = nodearray_definitions.get_machinetype(nodearray_name, resources.get("machinetype"), group_id)
else:
# rely solely on resources_available
pbscc.debug("machinetype is not defined for host %s, relying only on resources_available" % hostname)
machinetype = {"availableCount": 1, "name": "undefined"}
inst = machine.new_machine_instance(machinetype, **pbsnode["resources_available"])
return inst
def autoscale(self):
'''
The main loop described at the top of this class.
Returns machine_requests, idle_machines and total_machines for ease of unit testing.
'''
pbscc.info("Begin autoscale cycle")
nodearray_definitions = self.fetch_nodearray_definitions()
pbsnodes_by_hostname, existing_machines, booting_instance_ids, instance_ids_to_shutdown = self.get_existing_machines(nodearray_definitions)
start_enabled = "true" == str(self.cc_config.get("cyclecloud.cluster.autoscale.start_enabled", "true")).lower()
if not start_enabled:
pbscc.warn("cyclecloud.cluster.autoscale.start_enabled is false, new machines will not be allocated.")
autoscaler = autoscalerlib.Autoscaler(nodearray_definitions, existing_machines, self.default_placement_attrs, start_enabled)
# throttle how many jobs we attempt to match. When pbspro.compress_jobs is true (default) this shouldn't really be an issue
# unless the user has over $pbspro.max_unmatched_jobs unique sets of requirements.
max_unmatched_jobs = int(self.cc_config.get("pbspro.max_unmatched_jobs", 10000))
unmatched_jobs = 0
for job in self.query_jobs():
if job.executing_hostname:
try:
autoscaler.get_machine(hostname=job.executing_hostname).add_job(job, force=True)
continue
except RuntimeError as e:
pbscc.error(str(e))
pass
if not autoscaler.add_job(job):
unmatched_jobs += 1
pbscc.info("Can not match job %s." % job.name)
if max_unmatched_jobs > 0 and unmatched_jobs >= max_unmatched_jobs:
pbscc.warn('Maximum number of unmatched jobs reached - %s. To configure this setting, change {"pbspro": "max_unmatched_jobs": N}} in %s' % (unmatched_jobs, pbscc.CONFIG_PATH))
break
machine_requests = autoscaler.get_new_machine_requests()
idle_machines = autoscaler.get_idle_machines()
autoscale_request = autoscale_util.create_autoscale_request(machine_requests)
for request_set in autoscale_request["sets"]:
configuration = request_set["nodeAttributes"]["Configuration"]
if "pbspro" not in configuration:
configuration["pbspro"] = {}
configuration["pbspro"]["slot_type"] = request_set["nodearray"]
if not request_set.get("placementGroupId"):
configuration["pbspro"]["is_grouped"] = False
else:
configuration["pbspro"]["is_grouped"] = True
autoscale_util.scale_up(self.clusters_api, autoscale_request)
for r in machine_requests:
if r.placeby_value:
pbscc.info("Requesting %d %s machines in placement group %s for nodearray %s" % (r.instancecount, r.machinetype, r.placeby_value, r.nodearray))
else:
pbscc.info("Requesting %d %s machines in nodearray %s" % (r.instancecount, r.machinetype, r.nodearray))
if pbscc.is_fine():
pbscc.fine("New target state of the cluster, including booting nodes:")
for m in autoscaler.machines:
pbscc.fine(" %s" % str(m))
if instance_ids_to_shutdown:
pbscc.info("Shutting down instance ids %s" % instance_ids_to_shutdown.keys())
self.clusters_api.shutdown(instance_ids_to_shutdown.keys())
for hostname in instance_ids_to_shutdown.itervalues():
pbscc.info("Deleting %s" % hostname)
self.driver.delete_host(hostname)
now = self.clock.time()
stop_enabled = "true" == str(self.cc_config.get("cyclecloud.cluster.autoscale.stop_enabled", "true")).lower()
if not stop_enabled:
pbscc.warn("cyclecloud.cluster.autoscale.stop_enabled is false, idle machines will not be terminated")
if stop_enabled:
idle_before_threshold = float(self.cc_config.get("cyclecloud.cluster.autoscale.idle_time_before_jobs", 3600))
idle_after_threshold = float(self.cc_config.get("cyclecloud.cluster.autoscale.idle_time_after_jobs", 300))
for m in idle_machines:
if m.get_attr("instance_id", "") not in booting_instance_ids:
pbscc.debug("Could not find instance id in CycleCloud %s" % m.get_attr("instance_id", ""))
continue
pbsnode = pbsnodes_by_hostname.get(m.hostname)
# the machine may not have converged yet, so
if pbsnode:
if "busy" in pbsnode["state"]:
if "down" in pbsnode["state"]:
pbscc.warn("WARNING: %s is down but busy with jobs %s", m.hostname, pbsnode.get("jobs", []))
else:
pbscc.error("WARNING: Falsely determined that %s is idle!" % m.hostname)
continue
last_state_change_time = pbsnode["last_state_change_time"]
last_used_time = pbsnode.get("last_used_time")
if last_used_time:
# last_used_time can be stale while a job is exiting, e.g. last_state_change_time could be < 5 minutes but
# somehow last_used_time > 5 minutes, causing us to prematurely terminate the node just because a job took a long time
# to exit.
last_used_time = max(last_state_change_time, last_used_time)
else:
last_used_time = self.clock.time()
if now - last_used_time > idle_after_threshold:
pbscc.info("Setting %s offline after %s seconds" % (m.hostname, now - last_used_time))
self.driver.set_offline(m.hostname)
elif now - last_state_change_time > idle_before_threshold:
pbscc.info("Setting %s offline after %s seconds" % (m.hostname, now - last_state_change_time))
self.driver.set_offline(m.hostname)
pbscc.info("End autoscale cycle")
# returned for testing purposes
return machine_requests, idle_machines, autoscaler.machines