Files
ray-project--ray/python/ray/train/v2/tests/test_scheduling.py
T
2026-07-13 13:17:40 +08:00

142 lines
4.1 KiB
Python

from collections import defaultdict
import pytest
import ray
from ray.cluster_utils import Cluster
from ray.train import ScalingConfig
from ray.train.v2._internal.constants import WORKER_GROUP_START_TIMEOUT_S_ENV_VAR
from ray.train.v2.api.data_parallel_trainer import DataParallelTrainer
@pytest.fixture
def multi_cpu_node_cluster():
"""Yields a CPU cluster with 4 CPU nodes (4x4 CPU)."""
cluster = Cluster()
for _ in range(4):
cluster.add_node(num_cpus=4)
cluster.wait_for_nodes()
cluster.connect()
yield cluster
ray.shutdown()
cluster.shutdown()
@pytest.fixture
def multi_cpu_node_labeled_cluster():
cluster = Cluster()
for _ in range(2):
cluster.add_node(num_cpus=4, labels={"subcluster": "my_subcluster"})
for _ in range(2):
cluster.add_node(num_cpus=4, labels={"subcluster": "other_subcluster"})
cluster.wait_for_nodes()
cluster.connect()
yield cluster
ray.shutdown()
cluster.shutdown()
@pytest.mark.parametrize("placement_strategy", ["PACK", "SPREAD"])
def test_placement_strategy(multi_cpu_node_cluster, tmp_path, placement_strategy):
"""Tests the placement strategy of the worker group.
Nodes in this test are virtual but still have unique `node_id`s.
"""
scaling_config = ScalingConfig(
num_workers=4,
resources_per_worker={"CPU": 1},
placement_strategy=placement_strategy,
)
def train_fn(config):
node_id = ray.get_runtime_context().get_node_id()
(tmp_path / node_id).touch()
trainer = DataParallelTrainer(train_fn, scaling_config=scaling_config)
trainer.fit()
if placement_strategy == "PACK":
assert len(list(tmp_path.iterdir())) == 1
elif placement_strategy == "SPREAD":
assert len(list(tmp_path.iterdir())) == 4
@pytest.mark.parametrize("placement_strategy", ["STRICT_PACK", "STRICT_SPREAD"])
def test_infeasible_placement_strategy(
multi_cpu_node_cluster, tmp_path, monkeypatch, placement_strategy
):
"""Infeasible requests will continue retrying until the resources are available."""
@ray.remote(num_cpus=0)
def run():
monkeypatch.setenv(WORKER_GROUP_START_TIMEOUT_S_ENV_VAR, "2")
scaling_config = ScalingConfig(
num_workers=8,
resources_per_worker={"CPU": 1},
placement_strategy=placement_strategy,
)
trainer = DataParallelTrainer(lambda: None, scaling_config=scaling_config)
trainer.fit()
future = run.remote()
ready, _ = ray.wait([future], timeout=0.5)
assert not ready
# Add 4 more 8-CPU nodes to the cluster to make the request feasible.
for _ in range(4):
multi_cpu_node_cluster.add_node(num_cpus=8)
ray.get(future)
@pytest.mark.parametrize(
"label_selector, expected_subcluster_counts",
[
({"subcluster": "my_subcluster"}, {"my_subcluster": 2}),
(
[{"subcluster": "my_subcluster"}, {"subcluster": "my_subcluster"}],
{"my_subcluster": 2},
),
(
[{"subcluster": "my_subcluster"}, {"subcluster": "other_subcluster"}],
{"my_subcluster": 1, "other_subcluster": 1},
),
],
)
def test_label_selector(
multi_cpu_node_labeled_cluster, label_selector, expected_subcluster_counts
):
@ray.remote
class Counter:
def __init__(self):
self.subcluster_counts = defaultdict(int)
def increment(self, subcluster):
self.subcluster_counts[subcluster] += 1
def get(self):
return self.subcluster_counts
counter = Counter.remote()
scaling_config = ScalingConfig(
num_workers=2,
label_selector=label_selector,
)
def train_fn():
ray.get(
counter.increment.remote(
ray.get_runtime_context().get_node_labels()["subcluster"]
)
)
trainer = DataParallelTrainer(train_fn, scaling_config=scaling_config)
trainer.fit()
assert ray.get(counter.get.remote()) == expected_subcluster_counts
if __name__ == "__main__":
import sys
sys.exit(pytest.main(["-v", "-x", __file__]))