def run(self):
counter = 0
while True:
time.sleep(1.0)
replies = {}
try:
for node in self.nodes:
node_id = node["NodeID"]
stub = self.stubs[node_id]
reply = stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest(
include_memory_info=self.include_memory_info),
timeout=2)
reply_dict = MessageToDict(reply)
reply_dict["nodeId"] = node_id
replies[node["NodeManagerAddress"]] = reply_dict
with self._raylet_stats_lock:
for address, reply_dict in replies.items():
self._raylet_stats[address] = reply_dict
except Exception:
logger.exception(traceback.format_exc())
finally:
counter += 1
# From time to time, check if new nodes have joined the cluster
# and update self.nodes
if counter % 10:
self._update_nodes()
def node_stats(node_manager_address=None,
node_manager_port=None,
include_memory_info=True):
"""Returns NodeStats object describing memory usage in the cluster."""
import grpc
from ray.core.generated import node_manager_pb2
from ray.core.generated import node_manager_pb2_grpc
# We can ask any Raylet for the global memory info.
assert (node_manager_address is not None and node_manager_port is not None)
raylet_address = "{}:{}".format(node_manager_address, node_manager_port)
channel = grpc.insecure_channel(
raylet_address,
options=[
("grpc.max_send_message_length", MAX_MESSAGE_LENGTH),
("grpc.max_receive_message_length", MAX_MESSAGE_LENGTH),
],
)
stub = node_manager_pb2_grpc.NodeManagerServiceStub(channel)
node_stats = stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest(
include_memory_info=include_memory_info),
timeout=30.0)
return node_stats
def try_get_node_stats(num_retry=5, timeout=2):
reply = None
for _ in range(num_retry):
try:
reply = stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest(), timeout=timeout)
break
except grpc.RpcError:
continue
assert reply is not None
return reply
def get_workers():
raylet = ray.nodes()[0]
raylet_address = "{}:{}".format(raylet["NodeManagerAddress"],
raylet["NodeManagerPort"])
channel = grpc.insecure_channel(raylet_address)
stub = node_manager_pb2_grpc.NodeManagerServiceStub(channel)
return [
worker for worker in stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest()).core_workers_stats
if worker.worker_type != common_pb2.DRIVER
]
def get_num_workers():
raylet = ray.nodes()[0]
raylet_address = "{}:{}".format(raylet["NodeManagerAddress"],
raylet["NodeManagerPort"])
channel = grpc.insecure_channel(raylet_address)
stub = node_manager_pb2_grpc.NodeManagerServiceStub(channel)
return len([
worker for worker in stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest()).workers_stats
if not worker.is_driver
])
async def _update_node_stats(self):
for ip, stub in self._stubs.items():
node_info = DataSource.nodes.get(ip)
if node_info["state"] != "ALIVE":
continue
try:
reply = await stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest(), timeout=2)
reply_dict = node_stats_to_dict(reply)
DataSource.node_stats[ip] = reply_dict
except Exception:
logger.exception(f"Error updating node stats of {ip}.")
async def _update_node_stats(self):
for node_id, stub in self._stubs.items():
node_info = DataSource.nodes.get(node_id)
if node_info["state"] != "ALIVE":
continue
try:
reply = await stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest(
include_memory_info=self._collect_memory_info),
timeout=2)
reply_dict = node_stats_to_dict(reply)
DataSource.node_stats[node_id] = reply_dict
except Exception:
logger.exception(f"Error updating node stats of {node_id}.")
def test_initial_workers(shutdown_only):
# `num_cpus` should be <=2 because a Travis CI machine only has 2 CPU cores
ray.init(num_cpus=1,
include_dashboard=True,
_internal_config=json.dumps({"enable_multi_tenancy": True}))
raylet = ray.nodes()[0]
raylet_address = "{}:{}".format(raylet["NodeManagerAddress"],
raylet["NodeManagerPort"])
channel = grpc.insecure_channel(raylet_address)
stub = node_manager_pb2_grpc.NodeManagerServiceStub(channel)
wait_for_condition(lambda: len([
worker for worker in stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest()).workers_stats
if not worker.is_driver
]) == 1,
timeout=10)
def run(self):
counter = 0
while True:
time.sleep(1.0)
replies = {}
for node in self.nodes:
node_id = node["NodeID"]
stub = self.stubs[node_id]
reply = stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest(), timeout=2)
replies[node["NodeManagerAddress"]] = reply
with self._raylet_stats_lock:
for address, reply in replies.items():
self._raylet_stats[address] = MessageToDict(reply)
counter += 1
# From time to time, check if new nodes have joined the cluster
# and update self.nodes
if counter % 10:
self.update_nodes()
def stat(address):
if not address:
address = services.find_redis_address_or_die()
logger.info("Connecting to Ray instance at {}.".format(address))
ray.init(address=address)
import grpc
from ray.core.generated import node_manager_pb2
from ray.core.generated import node_manager_pb2_grpc
for raylet in ray.nodes():
raylet_address = "{}:{}".format(raylet["NodeManagerAddress"],
ray.nodes()[0]["NodeManagerPort"])
logger.info("Querying raylet {}".format(raylet_address))
channel = grpc.insecure_channel(raylet_address)
stub = node_manager_pb2_grpc.NodeManagerServiceStub(channel)
reply = stub.GetNodeStats(node_manager_pb2.GetNodeStatsRequest(),
timeout=2.0)
print(reply)
def stat(address):
"""Get the current metrics protobuf from a Ray cluster (developer tool)."""
if not address:
address = services.find_redis_address_or_die()
logger.info("Connecting to Ray instance at {}.".format(address))
ray.init(address=address)
import grpc
from ray.core.generated import node_manager_pb2
from ray.core.generated import node_manager_pb2_grpc
for raylet in ray.nodes():
raylet_address = "{}:{}".format(raylet["NodeManagerAddress"],
ray.nodes()[0]["NodeManagerPort"])
logger.info("Querying raylet {}".format(raylet_address))
channel = grpc.insecure_channel(raylet_address)
stub = node_manager_pb2_grpc.NodeManagerServiceStub(channel)
reply = stub.GetNodeStats(
node_manager_pb2.GetNodeStatsRequest(include_memory_info=False),
timeout=2.0)
print(reply)
