ray/python/ray/tests/test_object_manager.py

from collections import defaultdict
import multiprocessing
import numpy as np
import pytest
import time
import warnings

import ray
from ray.cluster_utils import Cluster
from ray.exceptions import GetTimeoutError

if (multiprocessing.cpu_count() < 40
        or ray.utils.get_system_memory() < 50 * 10**9):
    warnings.warn("This test must be run on large machines.")


def create_cluster(num_nodes):
    cluster = Cluster()
    for i in range(num_nodes):
        cluster.add_node(resources={str(i): 100}, object_store_memory=10**9)

    ray.init(address=cluster.address)
    return cluster


@pytest.fixture()
def ray_start_cluster_with_resource():
    num_nodes = 5
    cluster = create_cluster(num_nodes)
    yield cluster, num_nodes

    # The code after the yield will run as teardown code.
    ray.shutdown()
    cluster.shutdown()


@pytest.mark.parametrize(
    "ray_start_cluster_head", [{
        "num_cpus": 0,
        "object_store_memory": 75 * 1024 * 1024,
    }],
    indirect=True)
def test_object_transfer_during_oom(ray_start_cluster_head):
    cluster = ray_start_cluster_head
    cluster.add_node(object_store_memory=75 * 1024 * 1024)

    @ray.remote
    def put():
        return np.random.rand(5 * 1024 * 1024)  # 40 MB data

    local_ref = ray.put(np.random.rand(5 * 1024 * 1024))
    remote_ref = put.remote()

    with pytest.raises(GetTimeoutError):
        ray.get(remote_ref, timeout=1)
    del local_ref
    ray.get(remote_ref)


# This test is here to make sure that when we broadcast an object to a bunch of
# machines, we don't have too many excess object transfers.
@pytest.mark.skip(reason="TODO(ekl)")
def test_object_broadcast(ray_start_cluster_with_resource):
    cluster, num_nodes = ray_start_cluster_with_resource

    @ray.remote
    def f(x):
        return

    x = np.zeros(1024 * 1024, dtype=np.uint8)

    @ray.remote
    def create_object():
        return np.zeros(1024 * 1024, dtype=np.uint8)

    object_refs = []

    for _ in range(3):
        # Broadcast an object to all machines.
        x_id = ray.put(x)
        object_refs.append(x_id)
        ray.get([
            f._remote(args=[x_id], resources={str(i % num_nodes): 1})
            for i in range(10 * num_nodes)
        ])

    for _ in range(3):
        # Broadcast an object to all machines.
        x_id = create_object.remote()
        object_refs.append(x_id)
        ray.get([
            f._remote(args=[x_id], resources={str(i % num_nodes): 1})
            for i in range(10 * num_nodes)
        ])

    # Wait for profiling information to be pushed to the profile table.
    time.sleep(1)
    transfer_events = ray.object_transfer_timeline()

    # Make sure that each object was transferred a reasonable number of times.
    for x_id in object_refs:
        relevant_events = [
            event for event in transfer_events
            if event["cat"] == "transfer_send"
            and event["args"][0] == x_id.hex() and event["args"][2] == 1
        ]

        # NOTE: Each event currently appears twice because we duplicate the
        # send and receive boxes to underline them with a box (black if it is a
        # send and gray if it is a receive). So we need to remove these extra
        # boxes here.
        deduplicated_relevant_events = [
            event for event in relevant_events if event["cname"] != "black"
        ]
        assert len(deduplicated_relevant_events) * 2 == len(relevant_events)
        relevant_events = deduplicated_relevant_events

        # Each object must have been broadcast to each remote machine.
        assert len(relevant_events) >= num_nodes - 1
        # If more object transfers than necessary have been done, print a
        # warning.
        if len(relevant_events) > num_nodes - 1:
            warnings.warn("This object was transferred {} times, when only {} "
                          "transfers were required.".format(
                              len(relevant_events), num_nodes - 1))
        # Each object should not have been broadcast more than once from every
        # machine to every other machine. Also, a pair of machines should not
        # both have sent the object to each other.
        assert len(relevant_events) <= (num_nodes - 1) * num_nodes / 2

        # Make sure that no object was sent multiple times between the same
        # pair of object managers.
        send_counts = defaultdict(int)
        for event in relevant_events:
            # The pid identifies the sender and the tid identifies the
            # receiver.
            send_counts[(event["pid"], event["tid"])] += 1
        assert all(value == 1 for value in send_counts.values())


# When submitting an actor method, we try to pre-emptively push its arguments
# to the actor's object manager. However, in the past we did not deduplicate
# the pushes and so the same object could get shipped to the same object
# manager many times. This test checks that that isn't happening.
def test_actor_broadcast(ray_start_cluster_with_resource):
    cluster, num_nodes = ray_start_cluster_with_resource

    @ray.remote
    class Actor:
        def ready(self):
            pass

        def set_weights(self, x):
            pass

    actors = [
        Actor._remote(
            args=[],
            kwargs={},
            num_cpus=0.01,
            resources={str(i % num_nodes): 1}) for i in range(30)
    ]

    # Wait for the actors to start up.
    ray.get([a.ready.remote() for a in actors])

    object_refs = []

    # Broadcast a large object to all actors.
    for _ in range(5):
        x_id = ray.put(np.zeros(1024 * 1024, dtype=np.uint8))
        object_refs.append(x_id)
        # Pass the object into a method for every actor.
        ray.get([a.set_weights.remote(x_id) for a in actors])

    # Wait for profiling information to be pushed to the profile table.
    time.sleep(1)
    transfer_events = ray.object_transfer_timeline()

    # Make sure that each object was transferred a reasonable number of times.
    for x_id in object_refs:
        relevant_events = [
            event for event in transfer_events if
            event["cat"] == "transfer_send" and event["args"][0] == x_id.hex()
        ]

        # NOTE: Each event currently appears twice because we duplicate the
        # send and receive boxes to underline them with a box (black if it is a
        # send and gray if it is a receive). So we need to remove these extra
        # boxes here.
        deduplicated_relevant_events = [
            event for event in relevant_events if event["cname"] != "black"
        ]
        assert len(deduplicated_relevant_events) * 2 == len(relevant_events)
        relevant_events = deduplicated_relevant_events

        # Each object must have been broadcast to each remote machine.
        assert len(relevant_events) >= num_nodes - 1
        # If more object transfers than necessary have been done, print a
        # warning.
        if len(relevant_events) > num_nodes - 1:
            warnings.warn("This object was transferred {} times, when only {} "
                          "transfers were required.".format(
                              len(relevant_events), num_nodes - 1))
        # Each object should not have been broadcast more than once from every
        # machine to every other machine. Also, a pair of machines should not
        # both have sent the object to each other.
        assert len(relevant_events) <= (num_nodes - 1) * num_nodes / 2

        # Make sure that no object was sent multiple times between the same
        # pair of object managers.
        send_counts = defaultdict(int)
        for event in relevant_events:
            # The pid identifies the sender and the tid identifies the
            # receiver.
            send_counts[(event["pid"], event["tid"])] += 1
        assert all(value == 1 for value in send_counts.values())


# The purpose of this test is to make sure we can transfer many objects. In the
# past, this has caused failures in which object managers create too many open
# files and run out of resources.
def test_many_small_transfers(ray_start_cluster_with_resource):
    cluster, num_nodes = ray_start_cluster_with_resource

    @ray.remote
    def f(*args):
        pass

    # This function creates 1000 objects on each machine and then transfers
    # each object to every other machine.
    def do_transfers():
        id_lists = []
        for i in range(num_nodes):
            id_lists.append([
                f._remote(args=[], kwargs={}, resources={str(i): 1})
                for _ in range(1000)
            ])
        ids = []
        for i in range(num_nodes):
            for j in range(num_nodes):
                if i == j:
                    continue
                ids.append(
                    f._remote(
                        args=id_lists[j], kwargs={}, resources={str(i): 1}))

        # Wait for all of the transfers to finish.
        ray.get(ids)

    do_transfers()
    do_transfers()
    do_transfers()
    do_transfers()


# This is a basic test to ensure that the pull request retry timer is
# integrated properly. To test it, we create a 2 node cluster then do the
# following:
# (1) Fill up the driver's object store.
# (2) Fill up the remote node's object store.
# (3) Try to get the remote object. This should fail due to an OOM error caused
#     by step 1.
# (4) Allow the local object to be evicted.
# (5) Try to get the object again. Now the retry timer should kick in and
#     successfuly pull the remote object.
@pytest.mark.timeout(30)
def test_pull_request_retry(shutdown_only):
    cluster = Cluster()
    cluster.add_node(num_cpus=0, num_gpus=1, object_store_memory=100 * 2**20)
    cluster.add_node(num_cpus=1, num_gpus=0, object_store_memory=100 * 2**20)
    cluster.wait_for_nodes()
    ray.init(address=cluster.address)

    @ray.remote
    def put():
        return np.zeros(64 * 2**20, dtype=np.int8)

    @ray.remote(num_cpus=0, num_gpus=1)
    def driver():
        local_ref = ray.put(np.zeros(64 * 2**20, dtype=np.int8))

        remote_ref = put.remote()

        ready, _ = ray.wait([remote_ref], timeout=1)
        assert len(ready) == 0

        del local_ref

        # This should always complete within 10 seconds.
        ready, _ = ray.wait([remote_ref], timeout=20)
        assert len(ready) > 0

    # Pretend the GPU node is the driver. We do this to force the placement of
    # the driver and `put` task on different nodes.
    ray.get(driver.remote())


@pytest.mark.timeout(30)
def test_pull_bundles_admission_control(shutdown_only):
    cluster = Cluster()
    object_size = int(6e6)
    num_objects = 10
    num_tasks = 10
    # Head node can fit all of the objects at once.
    cluster.add_node(
        num_cpus=0,
        object_store_memory=2 * num_tasks * num_objects * object_size)
    cluster.wait_for_nodes()
    ray.init(address=cluster.address)

    # Worker node can only fit 1 task at a time.
    cluster.add_node(
        num_cpus=1, object_store_memory=1.5 * num_objects * object_size)
    cluster.wait_for_nodes()

    @ray.remote
    def foo(*args):
        return

    args = []
    for _ in range(num_tasks):
        task_args = [
            ray.put(np.zeros(object_size, dtype=np.uint8))
            for _ in range(num_objects)
        ]
        args.append(task_args)

    tasks = [foo.remote(*task_args) for task_args in args]
    ray.get(tasks)


@pytest.mark.timeout(30)
def test_pull_bundles_admission_control_dynamic(shutdown_only):
    # This test is the same as test_pull_bundles_admission_control, except that
    # the object store's capacity starts off higher and is later consumed
    # dynamically by concurrent workers.
    cluster = Cluster()
    object_size = int(6e6)
    num_objects = 10
    num_tasks = 10
    # Head node can fit all of the objects at once.
    cluster.add_node(
        num_cpus=0,
        object_store_memory=2 * num_tasks * num_objects * object_size)
    cluster.wait_for_nodes()
    ray.init(address=cluster.address)

    # Worker node can fit 2 tasks at a time.
    cluster.add_node(
        num_cpus=1, object_store_memory=2.5 * num_objects * object_size)
    cluster.wait_for_nodes()

    @ray.remote
    def foo(i, *args):
        print("foo", i)
        return

    @ray.remote
    def allocate(i):
        print("allocate", i)
        return np.zeros(object_size, dtype=np.uint8)

    args = []
    for _ in range(num_tasks):
        task_args = [
            ray.put(np.zeros(object_size, dtype=np.uint8))
            for _ in range(num_objects)
        ]
        args.append(task_args)

    tasks = [foo.remote(i, *task_args) for i, task_args in enumerate(args)]
    allocated = [allocate.remote(i) for i in range(num_objects)]
    ray.get(tasks)
    del allocated


if __name__ == "__main__":
    import pytest
    import sys
    sys.exit(pytest.main(["-v", __file__]))