def fetch_nodearray_definitions(self):
'''
A wrapper around the autoscale library function to parse Configuration.autoscale.* chef attributes and add
the 'ungrouped' attribute to the machine types.
See cyclecloud.nodearrays.NodearrayDefinitions for more info.
'''
nodearray_definitions = machine.fetch_nodearray_definitions(self.clusters_api, self.default_placement_attrs)
nodearray_definitions.placement_group_optional = True
filtered_nodearray_definitions = nodearrays.NodearrayDefinitions()
for machinetype in nodearray_definitions:
if machinetype.get("disabled", False):
continue
# ensure that any custom attribute the user specified, like disk = 100G, gets parsed correctly
for key, value in machinetype.iteritems():
try:
machinetype[key] = pbscc.parse_gb_size(key, value)
except InvalidSizeExpressionError:
pass
# kludge: there is a strange bug where ungrouped is showing up as a string and not a boolean.
if not machinetype.get("group_id"):
machinetype["ungrouped"] = "true"
filtered_nodearray_definitions.add_machinetype(machinetype)
else:
machinetype["ungrouped"] = "false"
filtered_nodearray_definitions.add_machinetype_with_placement_group(machinetype.get("group_id"), machinetype)
return filtered_nodearray_definitions
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 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
