def query_jobs(self):
'''
Converts PBS jobs into cyclecloud.job.Job instances. It will also compress jobs that have the exact same requirements.
'''
scheduler_config = self.driver.scheduler_config()
scheduler_resources = [] + scheduler_config["resources"]
# special case for hostname so we can automatically place jobs onto the appropriate host
scheduler_resources.append("hostname")
scheduler_resources.append("instance_id")
group_jobs = not self.disable_grouping
running_autoscale_jobs = []
idle_autoscale_jobs = []
# get the raw string outputs first, and convert it second. This somewhat limits the
# race condition of asking the status of the queue twice.
running_raw_jobs_str, running_converter = self.driver.running_jobs()
queued_raw_jobs_str, queued_converter = self.driver.queued_jobs()
running_raw_jobs = running_converter(running_raw_jobs_str)
queued_raw_jobs = queued_converter(queued_raw_jobs_str)
raw_jobs = []
for raw_job in running_raw_jobs:
# it is only running on a single node
if '+' not in raw_job["exec_vnode"]:
raw_jobs.append(raw_job)
continue
for vnode in raw_job["exec_vnode"].split("+"):
sub_raw_job = deepcopy(raw_job)
sub_raw_job["exec_vnode"] = vnode
raw_jobs.append(sub_raw_job)
for raw_job in queued_raw_jobs:
if not raw_job["resource_list"].get("select"):
raw_jobs.append(raw_job)
else:
# pbspro, like many schedulers, allows a varying set requirements for nodes in a single submission.
# we will break it apart here as if they had split them individually.
place = raw_job["resource_list"].get("place")
slot_type = raw_job["resource_list"].get("slot_type")
chunks = pbscc.parse_select(raw_job)
for n, chunk in enumerate(chunks):
# only pay the penalty of copies when we actually have multi-chunk jobs
sub_raw_job = deepcopy(raw_job)
if len(chunks) > 1:
sub_raw_job["job_id"] = "%s.%d" % (sub_raw_job["job_id"], n)
sub_raw_job["resource_list"] = {}
if place:
sub_raw_job["resource_list"]["place"] = place
if slot_type:
sub_raw_job["resource_list"]["slot_type"] = slot_type
sub_raw_job["resource_list"]["select"] = pbscc.format_select(chunk)
chunk["nodect"] = int(chunk["select"])
if "ncpus" not in chunk:
chunk["ncpus"] = "1"
for key, value in chunk.iteritems():
if key not in ["select", "nodect"]:
try:
value = pbscc.parse_gb_size(key, value) * chunk["nodect"]
except InvalidSizeExpressionError:
pass
sub_raw_job["resource_list"][key] = value
sub_raw_job["nodect"] = sub_raw_job["resource_list"]["nodect"] = chunk["nodect"]
raw_jobs.append(sub_raw_job)
warnings = set()
raw_jobs = [x for x in raw_jobs if x["job_state"].upper() in [pbscc.JOB_STATE_QUEUED,
pbscc.JOB_STATE_RUNNING,
pbscc.JOB_STATE_BATCH]]
for raw_job in raw_jobs:
pbs_job = mockpbs.mock_job(raw_job)
nodect = int(pbs_job.Resource_List["nodect"])
if pbs_job["job_state"].upper() == pbscc.JOB_STATE_RUNNING:
# update running job
live_resources = pbscc.parse_exec_vnode(raw_job["exec_vnode"])
for key, value in live_resources.iteritems():
# live resources are calculated on a per node basis, but the Resource_List is based
# on a total basis.
# we will normalize this below
if isinstance(value, numbers.Number):
pbs_job.Resource_List[key] = value * nodect
else:
pbs_job.Resource_List[key] = value
pbs_job["executing_hostname"] = live_resources["hostname"]
is_array = bool(pbs_job.get("array", False))
slots_per_job = int(pbs_job.Resource_List['ncpus']) / nodect
slot_type = pbs_job.Resource_List["slot_type"] # can be None, similar to {}.get("key"). It is a pbs class.
pbscc.info("found slot_type %s." % slot_type)
placement = pbscc.parse_place(pbs_job.Resource_List.get("place"))
# Note: not sure we will ever support anything but group_id for autoscale purposes.
# User could pick, say, group=host, which implies an SMP job, not a parallel job.
if placement.get("grouping", "group=group_id") != "group=group_id":
placement.pop("grouping")
if placement.get("arrangement", "").lower() in ["scatter", "vscatter"]:
pack = "scatter"
else:
pack = "pack"
exclusive = placement.get("sharing", "").lower() in ["excl", "exclhost"]
# we may need to support sharing at some point, but it seems that we can ignore it for now.
_shared = placement.get("sharing") in ["sharing"]
placeby = placement.get("grouping")
autoscale_job = Job(name=pbs_job["job_id"],
nodearray=slot_type,
nodes=nodect,
packing_strategy=pack,
exclusive=exclusive,
resources={"ncpus": 0},
executing_hostname=pbs_job.get("executing_hostname"))
if placeby:
autoscale_job.placeby = placeby.split("=", 1)[-1]
if is_array:
array_count = 0
array_tasks = raw_job["array_state_count"]
# Only grab the first two array task states (queued and running)
for ajob in str(array_tasks).split(" ")[:2]:
array_count += int(ajob.split(":")[1])
# Multiply the number of cpus needed by number of tasks in the array
if array_count != 0:
slots_per_job *= array_count
else:
array_count = 1
# If it's an MPI job and grouping is enabled
# we want to use a grouped autoscale_job to get tightly coupled nodes
if group_jobs and placement.get("grouping"):
autoscale_job['grouped'] = True
autoscale_job["nodes"] *= array_count
autoscale_job.placeby_value = "single"
elif is_array:
autoscale_job["nodes"] *= array_count
autoscale_job.ncpus += slots_per_job
for attr, value in pbs_job.Resource_List.iteritems():
if attr not in scheduler_resources:
# if it isn't a scheduler level attribute, don't bother
# considering it for autoscale as the scheduler won't respect it either.
continue
try:
value = pbscc.parse_gb_size(attr, value)
value = value / nodect
except InvalidSizeExpressionError:
if value.lower() in ["true", "false"]:
value = value.lower() == "true"
autoscale_job.resources[attr] = value
if raw_job["job_state"] == pbscc.JOB_STATE_QUEUED:
idle_autoscale_jobs.append(autoscale_job)
else:
running_autoscale_jobs.append(autoscale_job)
for warning in warnings:
format_string, values = warning[0], warning[1:]
pbscc.error(format_string % values)
# leave an option for disabling this in case it causes issues.
if self.cc_config.get("pbspro.compress_jobs", False):
all_autoscale_jobs = running_autoscale_jobs + compress_queued_jobs(idle_autoscale_jobs)
else:
all_autoscale_jobs = running_autoscale_jobs + idle_autoscale_jobs
return all_autoscale_jobs
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