def do_allocate(
self,
node_mgr: nodemanager.NodeManager,
allow_existing: bool,
all_or_nothing: bool,
) -> AllocationResult:
if self.__node_count > 0:
return node_mgr.allocate(
self._constraints,
node_count=self.__node_count,
allow_existing=allow_existing,
all_or_nothing=self.__colocated,
assignment_id=self.name,
)
assert self.iterations_remaining > 0
result = node_mgr.allocate(
self._constraints,
slot_count=self.iterations_remaining,
allow_existing=allow_existing,
# colocated is always all or nothing
all_or_nothing=all_or_nothing or self.colocated,
assignment_id=self.name,
)
if result:
self.iterations_remaining -= result.total_slots
return result
def add_ccnodeid_default_resource(node_mgr: NodeManager) -> None:
"""
In order for this to function properly, ccnodeid must be defined
"""
def get_node_id(n: Node) -> Optional[str]:
return n.delayed_node_id.node_id
node_mgr.add_default_resource({}, "ccnodeid", get_node_id)
def test_top_level_limits(node_mgr: NodeManager) -> None:
assert node_mgr.cluster_max_core_count == 10_000
assert node_mgr.cluster_consumed_core_count == 0
assert ["westus2"] == node_mgr.get_locations()
assert node_mgr.get_regional_consumed_core_count("westus2") == 0
assert node_mgr.get_regional_max_core_count("westus2") == 80
assert node_mgr.allocate({"node.vcpu_count": 4}, node_count=1)
assert node_mgr.cluster_consumed_core_count == 4
assert node_mgr.get_regional_consumed_core_count("westus2") == 4
def test_over_allocate(node_mgr: NodeManager) -> None:
assert node_mgr.allocate({"node.nodearray": "htc"}, node_count=1)
# can't allocate 10, because there are only 9 left
assert not node_mgr.allocate(
{"node.nodearray": "htc"}, node_count=10, all_or_nothing=True)
result = node_mgr.allocate({"node.nodearray": "htc"},
node_count=10,
all_or_nothing=False)
assert result and len(result.nodes) == 9
assert result.nodes[0].nodearray == "htc"
def make_requests(node_mgr: NodeManager) -> None:
for n, mag in enumerate(magnitudes):
node_count = None if slots_or_nodes[n] else mag
slot_count = None if node_count else mag
node_mgr.allocate(
{
"ncpus": ncpus_per_job[n],
"exclusive": exclusivity[n]
},
node_count=node_count,
slot_count=slot_count,
)
def test_no_buckets():
node_mgr = NodeManager(MockClusterBinding(), [])
dc = DemandCalculator(
node_mgr, NullNodeHistory(), singleton_lock=util.NullSingletonLock()
)
result = dc._add_job(Job("1", {"ncpus": 2}))
assert not result
assert "NoBucketsDefined" == result.status
def preprocess_node_mgr(self, config: Dict, node_mgr: NodeManager) -> None:
"""
We add a default resource to map group_id to node.placement_group
"""
super().preprocess_node_mgr(config, node_mgr)
def group_id(node: Node) -> str:
return node.placement_group if node.placement_group else "_none_"
node_mgr.add_default_resource({},
"group_id",
group_id,
allow_none=False)
def ungrouped(node: Node) -> str:
return str(not bool(node.placement_group)).lower()
node_mgr.add_default_resource({}, "ungrouped", ungrouped)
def validate_hg_intersections(ge_env: GridEngineEnvironment,
node_mgr: NodeManager,
warn_function: WarnFunction) -> bool:
bucket_to_hgs: Dict[str, Set[str]] = {}
for bucket in node_mgr.get_buckets():
if bucket.bucket_id not in bucket_to_hgs:
bucket_to_hgs[str(bucket)] = set()
by_str = partition_single(node_mgr.get_buckets(), str)
for queue in ge_env.queues.values():
if not queue.autoscale_enabled:
continue
for hostgroup in queue.bound_hostgroups.values():
for bucket in node_mgr.get_buckets():
is_satisfied = True
for constraint in hostgroup.constraints:
result = constraint.satisfied_by_bucket(bucket)
if not result:
is_satisfied = False
break
if is_satisfied:
bucket_to_hgs[str(bucket)].add(hostgroup.name)
failure = False
for bkey, matches in bucket_to_hgs.items():
bucket = by_str[bkey]
if not matches:
warn_function(
"%s is not matched by any hostgroup. This is not an error.",
bucket,
)
elif len(matches) > 1:
# seq_no will be used to determine ties
if not ge_env.scheduler.sort_by_seqno:
warn_function(
"%s is matched by more than one hostgroup %s. This is not an error.",
bucket,
",".join(matches),
)
return failure
def add_default_placement_groups(config: Dict, node_mgr: NodeManager) -> None:
nas = config.get("nodearrays", {})
for name, child in nas.items():
if child.get("placement_groups"):
return
by_pg = partition(node_mgr.get_buckets(), lambda b:
(b.nodearray, b.placement_group))
by_na_vm = partition(node_mgr.get_buckets(), lambda b:
(b.nodearray, b.vm_size))
for key, buckets in by_na_vm.items():
nodearray, vm_size = key
non_pg_buckets = [b for b in buckets if not b.placement_group]
if not non_pg_buckets:
# hardcoded PlacementGroupId
logging.debug(
"Nodearray %s defines PlacementGroupId, so no additional " +
"placement groups will be created automatically.",
nodearray,
)
continue
bucket = non_pg_buckets[0]
if not bucket.supports_colocation:
continue
buf_size = int(
nas.get(nodearray, {}).get("generated_placement_group_buffer", 2))
buf_remaining = buf_size
pgi = 0
while buf_remaining > 0:
pg_name = ht.PlacementGroup("{}_pg{}".format(vm_size, pgi))
pg_key = (nodearray, pg_name)
if pg_key not in by_pg:
logging.fine("Adding placement group %s", pg_name)
node_mgr.add_placement_group(pg_name, bucket)
buf_remaining -= 1
pgi += 1
def test_packing(node_mgr: NodeManager) -> None:
# htc node can fit 4 ncpus, so only allocate one node
result = node_mgr.allocate({
"node.nodearray": "htc",
"ncpus": 1
},
slot_count=2)
assert result, str(result)
assert len(result.nodes) == 1, result.nodes
assert result.nodes[0].name == "htc-1"
assert result.nodes[0].resources["ncpus"] == 4
assert result.nodes[0].available["ncpus"] == 2, result.nodes[0].available[
"ncpus"]
assert len(node_mgr.new_nodes) == 1, len(node_mgr.new_nodes)
# htc node can fit 4 ncpus, but 2 are used up on the first node, so allocate a second
result = node_mgr.allocate({
"node.nodearray": "htc",
"ncpus": 1
},
slot_count=4)
assert result
assert len(result.nodes) == 2, result.nodes
assert result.nodes[0].name == "htc-1"
assert result.nodes[1].name == "htc-2"
assert len(node_mgr.new_nodes) == 2, [n.name for n in node_mgr.new_nodes]
assert len(set([n.name for n in node_mgr.new_nodes])) == 2
result = node_mgr.allocate({
"node.nodearray": "htc",
"ncpus": 1
},
slot_count=2)
assert len(result.nodes) == 1
assert result.nodes[0].name == "htc-2"
assert len(node_mgr.new_nodes) == 2
def test_single_alloc(node_mgr: NodeManager) -> None:
result = node_mgr.allocate({"node.nodearray": "htc"}, node_count=1)
assert result and len(result.nodes) == 1
assert result.nodes[0].nodearray == "htc"
def test_unmanaged_nodes(node_mgr: NodeManager) -> None:
assert len(node_mgr.get_buckets()) == 2
tux = SchedulerNode("tux", bucket_id=ht.BucketId("tuxid"))
node_mgr.add_unmanaged_nodes([tux])
assert len(node_mgr.get_buckets()) == 3
assert node_mgr.get_buckets_by_id()[tux.bucket_id].nodes == [tux]
tux2 = SchedulerNode("tux2", bucket_id=tux.bucket_id)
node_mgr.add_unmanaged_nodes([tux2])
assert len(node_mgr.get_buckets()) == 3
assert node_mgr.get_buckets_by_id()[tux.bucket_id].nodes == [tux, tux2]
node_mgr.add_unmanaged_nodes([tux, tux2])
assert len(node_mgr.get_buckets()) == 3
assert node_mgr.get_buckets_by_id()[tux.bucket_id].nodes == [tux, tux2]
def test_node_software_configuration_alias(node_mgr: NodeManager) -> None:
node_mgr.add_default_resource({}, "int_alias",
"node.software_configuration.custom_int")
b = node_mgr.get_buckets()[0]
assert b.resources["int_alias"] == b.software_configuration["custom_int"]
def test_node_resources_alias(node_mgr: NodeManager) -> None:
node_mgr.add_default_resource({}, "memgb_alias", "node.resources.memgb")
b = node_mgr.get_buckets()[0]
assert b.resources["memgb_alias"] == b.resources["memgb"]
