def testOthersTakingResources(self):
# Let someone occupy the head node
pg = placement_group([{"CPU": 4, "GPU": 1}])
ray.get(pg.ready())
# We are left with the second node
assert len(nodes()) == 1
assert default_device(refresh=True) == "GPU"
pg = placement_group([{"GPU": 1}])
ray.get(pg.ready())
# Default device should be CPU
assert default_device(refresh=True) == "CPU"
assert len(nodes()) == 1
def test_warning_for_infeasible_tasks(ray_start_regular, error_pubsub):
p = error_pubsub
# Check that we get warning messages for infeasible tasks.
@ray.remote(num_gpus=1)
def f():
pass
@ray.remote(resources={"Custom": 1})
class Foo:
pass
# This task is infeasible.
f.remote()
errors = get_error_message(p, 1, ray_constants.INFEASIBLE_TASK_ERROR)
assert len(errors) == 1
assert errors[0].type == ray_constants.INFEASIBLE_TASK_ERROR
# This actor placement task is infeasible.
Foo.remote()
errors = get_error_message(p, 1, ray_constants.INFEASIBLE_TASK_ERROR)
assert len(errors) == 1
assert errors[0].type == ray_constants.INFEASIBLE_TASK_ERROR
# Placement group cannot be made, but no warnings should occur.
pg = placement_group([{"GPU": 1}], strategy="STRICT_PACK")
pg.ready()
f.options(placement_group=pg).remote()
errors = get_error_message(
p, 1, ray_constants.INFEASIBLE_TASK_ERROR, timeout=5)
assert len(errors) == 0, errors
def _create_placement_group(num_cpus_per_actor, num_actors):
"""
Create Ray placement group to grab resources.
Parameters
----------
num_cpus_per_actor : int
Number of CPUs per actor.
num_actors : int
Number of actors.
Returns
-------
ray.util.PlacementGroup
Placement group with grabbed resources.
"""
cpu_bundle = {"CPU": num_cpus_per_actor}
bundles = [cpu_bundle for _ in range(num_actors)]
pg = placement_group(bundles, strategy="SPREAD")
ready, _ = ray.wait([pg.ready()], timeout=100)
if ready is None:
raise TimeoutError("Placement group creation timeout.")
return pg
def pg_launcher(num_pgs_to_create):
print("Creating pgs")
pgs = []
for i in range(num_pgs_to_create):
pgs.append(placement_group(bundles, strategy="STRICT_SPREAD"))
pgs_removed = []
pgs_unremoved = []
# Randomly choose placement groups to remove.
if pg_removal:
print("removing pgs")
for pg in pgs:
if random() < 0.5 and pg_removal:
pgs_removed.append(pg)
else:
pgs_unremoved.append(pg)
print(len(pgs_unremoved))
tasks = []
# Randomly schedule tasks or actors on placement groups that
# are not removed.
for pg in pgs_unremoved:
for i in range(num_nodes):
tasks.append(
mock_task.options(placement_group=pg,
placement_group_bundle_index=i).remote())
# Remove the rest of placement groups.
if pg_removal:
for pg in pgs_removed:
remove_placement_group(pg)
ray.get(tasks)
# Since placement groups are scheduled, remove them.
for pg in pgs_unremoved:
remove_placement_group(pg)
def main():
ray.init(address="auto")
bundles = [{"CPU": 1, "GPU": 1}]
bundles += [{"CPU": 1} for _ in range(NUM_CPU_BUNDLES)]
pg = placement_group(bundles, strategy="PACK")
ray.get(pg.ready())
workers = [
Worker.options(placement_group=pg).remote(i)
for i in range(NUM_CPU_BUNDLES)
]
trainer = Trainer.options(placement_group=pg).remote(0)
start = time.time()
while True:
ray.get([workers[i].work.remote() for i in range(NUM_CPU_BUNDLES)])
ray.get(trainer.train.remote())
end = time.time()
if end - start > RUNTIME:
break
if "TEST_OUTPUT_JSON" in os.environ:
out_file = open(os.environ["TEST_OUTPUT_JSON"], "w")
results = {}
json.dump(results, out_file)
def pg_launcher(pre_created_pgs, num_pgs_to_create):
pgs = []
pgs += pre_created_pgs
for i in range(num_pgs_to_create):
pgs.append(
placement_group(bundles, strategy="STRICT_SPREAD", name=str(i)))
pgs_removed = []
pgs_unremoved = []
# Randomly choose placement groups to remove.
for pg in pgs:
if random() < .5:
pgs_removed.append(pg)
else:
pgs_unremoved.append(pg)
# Randomly schedule tasks or actors on placement groups that
# are not removed.
for pg in pgs_unremoved:
# TODO(sang): Comment in this line causes GCS actor management
# failure. We need to fix it.
# if random() < .5:
mock_task.options(placement_group=pg).remote()
# else:
# MockActor.options(placement_group=pg).remote()
# Remove the rest of placement groups.
for pg in pgs_removed:
remove_placement_group(pg)
ray.get([pg.ready() for pg in pgs_unremoved], timeout=10)
# Since placement groups are scheduled, remove them.
for pg in pgs_unremoved:
remove_placement_group(pg)
def test_placement_group_removal_leak_regression(ray_start_cluster):
"""Related issue:
https://github.com/ray-project/ray/issues/19131
"""
cluster = ray_start_cluster
cluster.add_node(num_cpus=5)
ray.init(address=cluster.address)
TOTAL_CPUS = 8
bundles = [{"CPU": 1, "GPU": 1}]
bundles += [{"CPU": 1} for _ in range(TOTAL_CPUS - 1)]
pg = placement_group(bundles, strategy="PACK")
# Here, we simulate that the ready task is queued and
# the new node is up. As soon as the new node is up,
# the ready task is scheduled.
# See https://github.com/ray-project/ray/pull/19138
# for more details about the test.
o = pg.ready()
# Add an artificial delay until the new node is up.
time.sleep(3)
cluster.add_node(num_cpus=5, num_gpus=1)
ray.get(o)
bundle_resource_name = f"bundle_group_{pg.id.hex()}"
expected_bundle_wildcard_val = TOTAL_CPUS * 1000
# This should fail if there's a leakage
# because the bundle resources are never returned properly.
def check_bundle_leaks():
bundle_resources = ray.available_resources()[bundle_resource_name]
return expected_bundle_wildcard_val == bundle_resources
wait_for_condition(check_bundle_leaks)
def test_pg_actor_workloads(ray_start_regular_with_external_redis):
from ray.util.placement_group import placement_group
bundle1 = {"CPU": 1}
pg = placement_group([bundle1], strategy="STRICT_PACK")
ray.get(pg.ready())
@ray.remote
class Counter:
def r(self, v):
return v
def pid(self):
import os
return os.getpid()
c = Counter.options(placement_group=pg).remote()
r = ray.get(c.r.remote(10))
assert r == 10
print("GCS is killed")
pid = ray.get(c.pid.remote())
ray.worker._global_node.kill_gcs_server()
assert ray.get(c.r.remote(10)) == 10
ray.worker._global_node.start_gcs_server()
for _ in range(100):
assert pid == ray.get(c.pid.remote())
def spread_to_all_nodes(f: RemoteFunction):
nodes = ray.state.nodes()
resources = [{'CPU': f._num_cpus} for _ in range(len(nodes))]
pg = placement_group(resources, strategy="STRICT_SPREAD")
ray.get(pg.ready())
yield len(nodes), pg
remove_placement_group(pg)
def test_fractional_resources_handle_correct(ray_start_cluster):
cluster = ray_start_cluster
cluster.add_node(num_cpus=1000)
ray.init(address=cluster.address)
bundles = [{"CPU": 0.01} for _ in range(5)]
pg = placement_group(bundles, strategy="SPREAD")
ray.get(pg.ready(), timeout=10)
def test_placement_group_local_resource_view(monkeypatch, ray_start_cluster):
"""Please refer to https://github.com/ray-project/ray/pull/19911
for more details.
"""
with monkeypatch.context() as m:
# Increase broadcasting interval so that node resource will arrive
# at raylet after local resource all being allocated.
m.setenv("RAY_raylet_report_resources_period_milliseconds", "2000")
m.setenv("RAY_grpc_based_resource_broadcast", "true")
cluster = ray_start_cluster
cluster.add_node(num_cpus=16, object_store_memory=1e9)
cluster.wait_for_nodes()
# We need to init here so that we can make sure it's connecting to
# the raylet where it only has cpu resources.
# This is a hacky way to prevent scheduling hanging which will
# schedule <CPU:1> job to the node with GPU and for <GPU:1, CPU:1> task
# there is no node has this resource.
ray.init(address="auto")
cluster.add_node(num_cpus=16, num_gpus=1)
cluster.wait_for_nodes()
NUM_CPU_BUNDLES = 30
@ray.remote(num_cpus=1)
class Worker(object):
def __init__(self, i):
self.i = i
def work(self):
time.sleep(0.1)
print("work ", self.i)
@ray.remote(num_cpus=1, num_gpus=1)
class Trainer(object):
def __init__(self, i):
self.i = i
def train(self):
time.sleep(0.2)
print("train ", self.i)
bundles = [{"CPU": 1, "GPU": 1}]
bundles += [{"CPU": 1} for _ in range(NUM_CPU_BUNDLES)]
pg = placement_group(bundles, strategy="PACK")
ray.get(pg.ready())
# Local resource will be allocated and here we are to ensure
# local view is consistent and node resouce updates are discarded
workers = [
Worker.options(placement_group=pg).remote(i)
for i in range(NUM_CPU_BUNDLES)
]
trainer = Trainer.options(placement_group=pg).remote(0)
ray.get([workers[i].work.remote() for i in range(NUM_CPU_BUNDLES)])
ray.get(trainer.train.remote())
def _create_placement_group(num_cpus_per_actor, num_actors):
cpu_bundle = {"CPU": num_cpus_per_actor}
bundles = [cpu_bundle for _ in range(num_actors)]
pg = placement_group(bundles, strategy="SPREAD")
ready, _ = ray.wait([pg.ready()], timeout=100)
if ready is None:
raise TimeoutError("Placement group creation timeout.")
return pg
def main():
"""Run a long running placement group creation/removal tests.
This test runs 20 trials first and measure the P50 performance.
After that it runs trials for a long time and make sure the
P50 creation/scheduling/removal performance is not regressed
after the long running job.
"""
args, _ = parse_script_args()
NUM_PG_AT_EACH_STAGE = args.num_pgs_stage
NUM_PENDING_PG = args.num_pending_pgs
TOTAL_STAGE = args.num_stages
if args.local:
ray.init(resources={"custom": 100, "pending": 1})
else:
ray.init(address="auto")
assert ray.cluster_resources()["custom"] >= NUM_PG_AT_EACH_STAGE * 4
assert ray.cluster_resources()["pending"] >= 1
# Create pending placement groups.
pending_pgs = []
for _ in range(NUM_PENDING_PG):
# Right now, we don't have infeasible pgs,
# so this will simulate the pending pgs.
pending_pgs.append(placement_group([{"pending": 1}], strategy="PACK"))
(scheduling_perf, removing_perf,
creation_perf) = run_trial(20, NUM_PG_AT_EACH_STAGE)
(scheduling_perf_final, removing_perf_final,
creation_perf_final) = run_trial(TOTAL_STAGE, NUM_PG_AT_EACH_STAGE)
print(f"Scheduling performance 20 trials: {scheduling_perf}")
print(
f"Scheduling performance {TOTAL_STAGE} trials: {scheduling_perf_final}"
)
print(f"Removal performance 20 trials: {removing_perf}")
print(f"Removal performance {TOTAL_STAGE} trials: {removing_perf_final}")
print(f"Creation performance 20 trials: {creation_perf}")
print(f"Creation performance {TOTAL_STAGE} trials: {creation_perf_final}")
assert scheduling_perf["p50_ms"] * 100 > scheduling_perf_final["p50_ms"]
assert removing_perf["p50_ms"] * 100 > removing_perf_final["p50_ms"]
assert creation_perf["p50_ms"] * 100 > creation_perf_final["p50_ms"]
if "TEST_OUTPUT_JSON" in os.environ:
out_file = open(os.environ["TEST_OUTPUT_JSON"], "w")
results = {}
json.dump(results, out_file)
def test_warning_for_infeasible_tasks(ray_start_regular, error_pubsub):
p = error_pubsub
# Check that we get warning messages for infeasible tasks.
@ray.remote(num_gpus=1)
def f():
pass
@ray.remote(resources={"Custom": 1})
class Foo:
pass
# This task is infeasible.
f.remote()
errors = get_error_message(p, 1, ray_constants.INFEASIBLE_TASK_ERROR)
assert len(errors) == 1
assert errors[0].type == ray_constants.INFEASIBLE_TASK_ERROR
# This actor placement task is infeasible.
foo = Foo.remote()
print(foo)
errors = get_error_message(p, 1, ray_constants.INFEASIBLE_TASK_ERROR)
assert len(errors) == 1
assert errors[0].type == ray_constants.INFEASIBLE_TASK_ERROR
# Placement group cannot be made, but no warnings should occur.
total_cpus = ray.cluster_resources()["CPU"]
# Occupy one cpu by an actor
@ray.remote(num_cpus=1)
class A:
pass
a = A.remote()
print(a)
@ray.remote(num_cpus=total_cpus)
def g():
pass
pg = placement_group([{"CPU": total_cpus}], strategy="STRICT_PACK")
g.options(placement_group=pg).remote()
errors = get_error_message(p,
1,
ray_constants.INFEASIBLE_TASK_ERROR,
timeout=5)
assert len(errors) == 0, errors
def test_many_placement_groups():
@ray.remote(num_cpus=1, resources={"node": 0.02})
def f1():
sleep(10)
pass
@ray.remote(num_cpus=1)
def f2():
sleep(10)
pass
@ray.remote(resources={"node": 0.02})
def f3():
sleep(10)
pass
bundle1 = {"node": 0.02, "CPU": 1}
bundle2 = {"CPU": 1}
bundle3 = {"node": 0.02}
pgs = []
for _ in trange(MAX_PLACEMENT_GROUPS, desc="Creating pgs"):
pg = placement_group(bundles=[bundle1, bundle2, bundle3])
pgs.append(pg)
for pg in tqdm(pgs, desc="Waiting for pgs to be ready"):
ray.get(pg.ready())
refs = []
for pg in tqdm(pgs, desc="Scheduling tasks"):
ref1 = f1.options(placement_group=pg).remote()
ref2 = f2.options(placement_group=pg).remote()
ref3 = f3.options(placement_group=pg).remote()
refs.extend([ref1, ref2, ref3])
for _ in trange(10, desc="Waiting"):
sleep(1)
with tqdm() as p_bar:
while refs:
done, refs = ray.wait(refs)
p_bar.update()
for pg in tqdm(pgs, desc="Cleaning up pgs"):
remove_placement_group(pg)
def test_many_placement_groups():
# @ray.remote(num_cpus=1, resources={"node": 0.02})
@ray.remote
class C1:
def ping(self):
return "pong"
# @ray.remote(num_cpus=1)
@ray.remote
class C2:
def ping(self):
return "pong"
# @ray.remote(resources={"node": 0.02})
@ray.remote
class C3:
def ping(self):
return "pong"
bundle1 = {"node": 0.02, "CPU": 1}
bundle2 = {"CPU": 1}
bundle3 = {"node": 0.02}
pgs = []
for _ in tqdm.trange(MAX_PLACEMENT_GROUPS, desc="Creating pgs"):
pg = placement_group(bundles=[bundle1, bundle2, bundle3])
pgs.append(pg)
for pg in tqdm.tqdm(pgs, desc="Waiting for pgs to be ready"):
ray.get(pg.ready())
actors = []
for pg in tqdm.tqdm(pgs, desc="Scheduling tasks"):
actors.append(C1.options(placement_group=pg).remote())
actors.append(C2.options(placement_group=pg).remote())
actors.append(C3.options(placement_group=pg).remote())
not_ready = [actor.ping.remote() for actor in actors]
for _ in tqdm.trange(len(actors)):
ready, not_ready = ray.wait(not_ready)
assert ray.get(*ready) == "pong"
for pg in tqdm.tqdm(pgs, desc="Cleaning up pgs"):
remove_placement_group(pg)
def test_schedule_placement_groups_at_the_same_time():
ray.init(num_cpus=4)
pgs = [placement_group([{"CPU": 2}]) for _ in range(6)]
wait_pgs = {pg.ready(): pg for pg in pgs}
def is_all_placement_group_removed():
ready, _ = ray.wait(list(wait_pgs.keys()), timeout=0.5)
if ready:
ready_pg = wait_pgs[ready[0]]
remove_placement_group(ready_pg)
del wait_pgs[ready[0]]
if len(wait_pgs) == 0:
return True
return False
wait_for_condition(is_all_placement_group_removed)
def test_placement_group_gpu_unique_assigned(ray_start_cluster,
connect_to_client):
cluster = ray_start_cluster
cluster.add_node(num_gpus=4, num_cpus=4)
ray.init(address=cluster.address)
gpu_ids_res = set()
# Create placement group with 4 bundles using 1 GPU each.
num_gpus = 4
bundles = [{"GPU": 1, "CPU": 1} for _ in range(num_gpus)]
pg = placement_group(bundles)
ray.get(pg.ready())
# Actor using 1 GPU that has a method to get
# $CUDA_VISIBLE_DEVICES env variable.
@ray.remote(num_gpus=1, num_cpus=1)
class Actor:
def get_gpu(self):
import os
return os.environ["CUDA_VISIBLE_DEVICES"]
# Create actors out of order.
actors = []
actors.append(
Actor.options(placement_group=pg,
placement_group_bundle_index=0).remote())
actors.append(
Actor.options(placement_group=pg,
placement_group_bundle_index=3).remote())
actors.append(
Actor.options(placement_group=pg,
placement_group_bundle_index=2).remote())
actors.append(
Actor.options(placement_group=pg,
placement_group_bundle_index=1).remote())
for actor in actors:
gpu_ids = ray.get(actor.get_gpu.remote())
assert len(gpu_ids) == 1
gpu_ids_res.add(gpu_ids)
assert len(gpu_ids_res) == 4
def test_infeasible_pg(ray_start_cluster):
"""Test infeasible pgs are scheduled after new nodes are added."""
cluster = ray_start_cluster
cluster.add_node(num_cpus=2)
ray.init("auto")
bundle = {"CPU": 4, "GPU": 1}
pg = placement_group([bundle], name="worker_1", strategy="STRICT_PACK")
# Placement group is infeasible.
with pytest.raises(GetTimeoutError):
ray.get(pg.ready(), timeout=3)
state = ray.util.placement_group_table()[
pg.id.hex()]["stats"]["scheduling_state"]
assert state == "INFEASIBLE"
# Add a new node. PG can now be scheduled.
cluster.add_node(num_cpus=4, num_gpus=1)
assert ray.get(pg.ready(), timeout=10)
def pg_launcher(pre_created_pgs, num_pgs_to_create):
pgs = []
pgs += pre_created_pgs
for i in range(num_pgs_to_create):
pgs.append(placement_group(BUNDLES, strategy="STRICT_SPREAD"))
pgs_removed = []
pgs_unremoved = []
# Randomly choose placement groups to remove.
for pg in pgs:
if random() < .5:
pgs_removed.append(pg)
else:
pgs_unremoved.append(pg)
tasks = []
max_actor_cnt = 5
actor_cnt = 0
actors = []
# Randomly schedule tasks or actors on placement groups that
# are not removed.
for pg in pgs_unremoved:
# TODO(sang): Comment in this line causes GCS actor management
# failure. We need to fix it.
if random() < .5:
tasks.append(mock_task.options(placement_group=pg).remote())
else:
if actor_cnt < max_actor_cnt:
actors.append(MockActor.options(placement_group=pg).remote())
actor_cnt += 1
# Remove the rest of placement groups.
for pg in pgs_removed:
remove_placement_group(pg)
ray.get([pg.ready() for pg in pgs_unremoved])
ray.get(tasks)
ray.get([actor.ping.remote() for actor in actors])
# Since placement groups are scheduled, remove them.
for pg in pgs_unremoved:
remove_placement_group(pg)
def test_chaos_defer(monkeypatch, ray_start_cluster):
with monkeypatch.context() as m:
m.setenv("RAY_grpc_based_resource_broadcast", "true")
# defer for 3s
m.setenv(
"RAY_testing_asio_delay_us",
"NodeManagerService.grpc_client.PrepareBundleResources=2000000:2000000",
)
m.setenv("RAY_event_stats", "true")
cluster = ray_start_cluster
cluster.add_node(num_cpus=1, object_store_memory=1e9)
cluster.wait_for_nodes()
ray.init(address="auto") # this will connect to gpu nodes
cluster.add_node(num_cpus=0, num_gpus=1)
bundle = [{"GPU": 1}, {"CPU": 1}]
pg = placement_group(bundle)
# PG will not be ready within 3s
with pytest.raises(ray.exceptions.GetTimeoutError):
ray.get(pg.ready(), timeout=1)
# it'll be ready eventually
ray.get(pg.ready())
def __call__(self, *args, **kwargs):
kwargs.update(self._bound.kwargs)
# Call with bounded *args and **kwargs
return placement_group(*self._bound.args, **kwargs)
value = torch.randn(batch_size)
input = torch.randn(batch_size)
return {
"input": input,
"value": value,
"actor_id": self.id,
}
tot_gpus = 1
tot_cpus = 8
pg_cnt = min(int(tot_gpus / bundle1['GPU']), int(tot_cpus / bundle1['CPU']))
pgs = [
placement_group([bundle1], strategy="STRICT_PACK") for _ in range(pg_cnt)
]
# Wait until placement group is created.
ray.get([pg.ready() for pg in pgs])
for pg in pgs:
print(placement_group_table(pg))
# You can look at placement group states using this API.
print(placement_group_table(pg))
actors = [
AsyncActor.options(placement_group=pg).remote(i)
for i, pg in enumerate(pgs)
]
# Get a list of the IP addresses of the nodes that have joined the cluster.
import ray
from ray.util.placement_group import (
placement_group, )
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
# Two "CPU"s are available.
ray.init(num_cpus=2)
# Create a placement group.
pg = placement_group([{"CPU": 2}])
ray.get(pg.ready())
# Now, 2 CPUs are not available anymore because
# they are pre-reserved by the placement group.
@ray.remote(num_cpus=2)
def f():
return True
# Won't be scheduled because there are no 2 cpus.
f.remote()
# Will be scheduled because 2 cpus are reserved by the placement group.
f.options(scheduling_strategy=PlacementGroupSchedulingStrategy(
placement_group=pg)).remote()
time.sleep(5)
logger.info("Nodes have all joined. There are %s resources.",
ray.cluster_resources())
def hey(_):
time.sleep(0.01) # Sleep for 10ms
return b"hey"
num_connections = int(num_remote_cpus * 0.75)
num_threads = 2
time_to_run = "10s"
pg = placement_group(
[{
"CPU": 1
} for _ in range(expected_num_nodes)], strategy="STRICT_SPREAD")
ray.get(pg.ready())
# The number of replicas is the number of cores remaining after accounting
# for the one HTTP proxy actor on each node, the "hey" requester task on each
# node, and the serve controller.
# num_replicas = expected_num_nodes * (cpus_per_node - 2) - 1
num_replicas = ray.available_resources()["CPU"]
logger.info("Starting %i replicas", num_replicas)
client.create_backend(
"hey", hey, config=BackendConfig(num_replicas=num_replicas))
client.create_endpoint("hey", backend="hey", route="/hey")
@ray.remote
# In[ ]:
@ray.remote
def remote_fun(x):
return x
# In[ ]:
#tag::placement_group[]
# Create a placement group.
cpu_bundle = {"CPU": 3}
mini_cpu_bundle = {"CPU": 1}
pg = placement_group([cpu_bundle, mini_cpu_bundle])
ray.get(pg.ready())
print(placement_group_table(pg))
print(ray.available_resources())
# Run remote_fun in cpu_bundle
handle = remote_fun.options(placement_group=pg,
placement_group_bundle_index=0).remote(1)
#end::placement_group[]
# In[ ]:
#tag::runtime_env_local[]
@ray.remote(runtime_env=runtime_env)
def sup(x):
from bs4 import BeautifulSoup
def run_trial(total_stage, num_pg_per_stage):
creating_e2e_s = []
removing_e2e_s = []
# Create and remove placement groups.
for i in range(total_stage):
# Create pgs.
pgs = []
start = perf_counter()
for _ in range(num_pg_per_stage):
pgs.append(
placement_group(bundles=[{
"custom": 0.025
} for _ in range(4)],
strategy="PACK"))
logger.info(f"Created {num_pg_per_stage} pgs.")
ray.get([pg.ready() for pg in pgs])
end = perf_counter()
total_creating_time = end - start
logger.info(f"Creating {num_pg_per_stage} took "
f"{total_creating_time} seconds at stage {i}")
creating_e2e_s.append(total_creating_time * 1000.0)
# Remove pgs
start = perf_counter()
for _, pg in enumerate(pgs):
remove_placement_group(pg)
end = perf_counter()
total_removal_time = end - start
logger.info(f"removed {num_pg_per_stage} pgs took "
f"{total_removal_time} seconds at stage {i}")
removing_e2e_s.append(total_removal_time * 1000.0)
# time.sleep(1)
# Calculate the scheduling latency (excluding queueing time).
latencies = []
for entry in ray.util.placement_group_table().values():
latency = entry["stats"]["scheduling_latency_ms"]
latencies.append(latency)
latencies = sorted(latencies)
removing_e2e_s = sorted(removing_e2e_s)
creating_e2e_s = sorted(creating_e2e_s)
def get_scheduling_perf(latencies):
"""Return P10, 50, 95, 99 latency"""
p10 = latencies[int(len(latencies) * 0.1)]
p50 = latencies[int(len(latencies) * 0.5)]
p95 = latencies[int(len(latencies) * 0.95)]
p99 = latencies[int(len(latencies) * 0.99)]
return {"p10_ms": p10, "p50_ms": p50, "p95_ms": p95, "p99_ms": p99}
scheduling_perf = get_scheduling_perf(latencies)
removing_perf = get_scheduling_perf(removing_e2e_s)
creation_perf = get_scheduling_perf(creating_e2e_s)
wait_for_condition(
lambda: (ray.cluster_resources()["custom"] == ray.available_resources(
)["custom"]),
timeout=30,
)
wait_for_condition(
lambda: (ray.cluster_resources()["pending"] == ray.available_resources(
)["pending"]),
timeout=30,
)
return scheduling_perf, removing_perf, creation_perf
def verify_load_metrics(monitor, expected_resource_usage=None, timeout=30):
request_resources(num_cpus=42)
# add placement groups.
pg_demands = [{"GPU": 2}, {"extra_resource": 2}]
strategy = "STRICT_PACK"
pg = placement_group(pg_demands, strategy=strategy)
pg.ready()
time.sleep(2) # wait for placemnt groups to propogate.
# Disable event clearing for test.
monitor.event_summarizer.clear = lambda *a: None
visited_atleast_once = [set(), set()]
while True:
monitor.update_load_metrics()
monitor.update_resource_requests()
monitor.update_event_summary()
resource_usage = monitor.load_metrics._get_resource_usage()
# Check resource request propagation.
req = monitor.load_metrics.resource_requests
assert req == [{"CPU": 1}] * 42, req
pg_response_data = monitor.load_metrics.pending_placement_groups
assert_correct_pg(pg_response_data, pg_demands, strategy)
if "memory" in resource_usage[0]:
del resource_usage[0]["memory"]
visited_atleast_once[0].add("memory")
if "object_store_memory" in resource_usage[0]:
del resource_usage[0]["object_store_memory"]
visited_atleast_once[0].add("object_store_memory")
if "memory" in resource_usage[1]:
del resource_usage[1]["memory"]
visited_atleast_once[1].add("memory")
if "object_store_memory" in resource_usage[1]:
del resource_usage[1]["object_store_memory"]
visited_atleast_once[1].add("object_store_memory")
for key in list(resource_usage[0].keys()):
if key.startswith("node:"):
del resource_usage[0][key]
visited_atleast_once[0].add("node:")
for key in list(resource_usage[1].keys()):
if key.startswith("node:"):
del resource_usage[1][key]
visited_atleast_once[1].add("node:")
if expected_resource_usage is None:
if all(x for x in resource_usage[0:]):
break
elif all(x == y
for x, y in zip(resource_usage, expected_resource_usage)):
break
else:
timeout -= 1
time.sleep(1)
if timeout <= 0:
raise ValueError("Timeout. {} != {}".format(
resource_usage, expected_resource_usage))
# Sanity check we emitted a resize event.
assert any("Resized to" in x for x in monitor.event_summarizer.summary())
assert visited_atleast_once[0] == {
"memory", "object_store_memory", "node:"
}
assert visited_atleast_once[0] == visited_atleast_once[1]
remove_placement_group(pg)
return resource_usage
ray.cluster_resources())
# Scenario 1: Create bunch of placement groups and measure how long
# it takes.
total_creating_time = 0
total_removing_time = 0
repeat = 1
total_trial = repeat * NUM_PG
BUNDLES = [{"pg_custom": 1}] * NUM_NODES
# Create and remove placement groups.
for _ in range(repeat):
pgs = []
for i in range(NUM_PG):
start = perf_counter()
pgs.append(placement_group(BUNDLES, strategy="PACK"))
end = perf_counter()
logger.info(f"append_group iteration {i}")
total_creating_time += (end - start)
ray.get([pg.ready() for pg in pgs])
for i, pg in enumerate(pgs):
start = perf_counter()
remove_placement_group(pg)
end = perf_counter()
logger.info(f"remove_group iteration {i}")
total_removing_time += (end - start)
# Validate the correctness.
assert ray.cluster_resources()[
import time
import ray
from ray.util.placement_group import (placement_group, placement_group_table,
remove_placement_group)
if __name__ == "__main__":
ray.init(num_cpus=2, resources={"extra_resources": 2})
bundle_1 = {"CPU": 2}
bundle_2 = {"extra_resources": 2}
pg = placement_group([bundle_1, bundle_2], strategy="STRICT_PACK")
# You can also use ray.wait.
ready, unready = ray.wait([pg.ready()], timeout=5)
print(f"placement group status:{ready}")
print(placement_group_table(pg))
time.sleep(10)
