[Ray] Implement gc for ray task executor context (#3061)

Author: zhongchun

mars/deploy/oscar/tests/test_ray_dag.py

@@ -14,13 +14,16 @@
 
 import copy
 import os
+import time
 
 import pytest
 
+from .... import get_context
+from .... import tensor as mt
 from ....tests.core import DICT_NOT_EMPTY, require_ray
 from ....utils import lazy_import
 from ..local import new_cluster
-from ..session import new_session
+from ..session import new_session, get_default_async_session
 from ..tests import test_local
 from ..tests.session import new_test_session
 from ..tests.test_local import _cancel_when_tile, _cancel_when_execute
@@ -129,3 +132,38 @@ async def test_session_get_progress(ray_start_regular_shared2, create_cluster):
 @pytest.mark.parametrize("test_func", [_cancel_when_execute, _cancel_when_tile])
 def test_cancel(ray_start_regular_shared2, create_cluster, test_func):
     test_local.test_cancel(create_cluster, test_func)
+
+
+@require_ray
+@pytest.mark.parametrize("config", [{"backend": "ray"}])
+def test_executor_context_gc(config):
+    session = new_session(
+        backend=config["backend"],
+        n_cpu=2,
+        web=False,
+        use_uvloop=False,
+        config={"task.execution_config.ray.subtask_monitor_interval": 0},
+    )
+
+    assert session._session.client.web_address is None
+    assert session.get_web_endpoint() is None
+
+    def f1(c):
+        time.sleep(0.5)
+        return c
+
+    with session:
+        t1 = mt.random.randint(10, size=(100, 10), chunk_size=100)
+        t2 = mt.random.randint(10, size=(100, 10), chunk_size=50)
+        t3 = t2 + t1
+        t4 = t3.sum(0)
+        t5 = t4.map_chunk(f1)
+        r = t5.execute()
+        result = r.fetch()
+        assert result is not None
+        assert len(result) == 10
+        context = get_context()
+        assert len(context._task_context) < 5
+
+    session.stop_server()
+    assert get_default_async_session() is None

mars/lib/ordered_set.py

@@ -44,7 +44,7 @@
 
 # SetLike[T] is either a set of elements of type T, or a sequence, which
 # we will convert to an OrderedSet by adding its elements in order.
-SetLike = Union[AbstractSet[T], Sequence[T]]
+SetLike = Union[AbstractSet[T], Sequence[T], Iterable[T]]
 OrderedSetInitializer = Union[AbstractSet[T], Sequence[T], Iterable[T]]
 
 

mars/services/task/execution/ray/config.py

@@ -17,6 +17,9 @@
 from ..api import ExecutionConfig, register_config_cls
 from ..utils import get_band_resources_from_config
 
+# The default interval seconds to update progress and collect garbage.
+DEFAULT_SUBTASK_MONITOR_INTERVAL = 1
+
 
 @register_config_cls
 class RayExecutionConfig(ExecutionConfig):
@@ -55,3 +58,12 @@ def create_task_state_actor_as_needed(self):
         #   - False:
         #     Create RayTaskState actor in advance when the RayTaskExecutor is created.
         return self._ray_execution_config.get("create_task_state_actor_as_needed", True)
+
+    def get_subtask_monitor_interval(self):
+        """
+        The interval seconds for the monitor task to update progress and
+        collect garbage.
+        """
+        return self._ray_execution_config.get(
+            "subtask_monitor_interval", DEFAULT_SUBTASK_MONITOR_INTERVAL
+        )

mars/services/task/execution/ray/executor.py

@@ -28,6 +28,7 @@
     execute,
 )
 from .....lib.aio import alru_cache
+from .....lib.ordered_set import OrderedSet
 from .....resource import Resource
 from .....serialization import serialize, deserialize
 from .....typing import BandType
@@ -96,7 +97,7 @@ def _optimize_subtask_graph(subtask_graph):
 
 
 async def _cancel_ray_task(obj_ref, kill_timeout: int = 3):
-    ray.cancel(obj_ref, force=False)
+    await asyncio.to_thread(ray.cancel, obj_ref, force=False)
     try:
         await asyncio.to_thread(ray.get, obj_ref, timeout=kill_timeout)
     except ray.exceptions.TaskCancelledError:  # pragma: no cover
@@ -108,7 +109,7 @@ async def _cancel_ray_task(obj_ref, kill_timeout: int = 3):
             e,
             obj_ref,
         )
-        ray.cancel(obj_ref, force=True)
+        await asyncio.to_thread(ray.cancel, obj_ref, force=True)
 
 
 def execute_subtask(
@@ -183,17 +184,12 @@ def __init__(
 
         self._available_band_resources = None
 
-        # For progress
+        # For progress and task cancel
         self._pre_all_stages_progress = 0.0
-        self._pre_all_stages_tile_progress = 0
-        self._cur_stage_tile_progress = 0
-        self._cur_stage_output_object_refs = []
-        # This list records the output object ref number of subtasks, so with
-        # `self._cur_stage_output_object_refs` we can just call `ray.cancel`
-        # with one object ref to cancel a subtask instead of cancel all object
-        # refs. In this way we can reduce a lot of unnecessary calls of ray.
-        self._output_object_refs_nums = []
-        # For meta and data gc
+        self._pre_all_stages_tile_progress = 0.0
+        self._cur_stage_progress = 0.0
+        self._cur_stage_tile_progress = 0.0
+        self._cur_stage_first_output_object_ref_to_subtask = dict()
         self._execute_subtask_graph_aiotask = None
         self._cancelled = False
 
@@ -258,12 +254,12 @@ def destroy(self):
 
         self._available_band_resources = None
 
-        # For progress
-        self._pre_all_stages_progress = 1
-        self._pre_all_stages_tile_progress = 1
-        self._cur_stage_tile_progress = 1
-        self._cur_stage_output_object_refs = []
-        self._output_object_refs_nums = []
+        # For progress and task cancel
+        self._pre_all_stages_progress = 1.0
+        self._pre_all_stages_tile_progress = 1.0
+        self._cur_stage_progress = 1.0
+        self._cur_stage_tile_progress = 1.0
+        self._cur_stage_first_output_object_ref_to_subtask = dict()
         self._execute_subtask_graph_aiotask = None
         self._cancelled = None
 
@@ -318,7 +314,33 @@ async def execute_subtask_graph(
     ) -> Dict[Chunk, ExecutionChunkResult]:
         if self._cancelled is True:  # pragma: no cover
             raise asyncio.CancelledError()
+
+        def _on_monitor_task_done(fut):
+            # Print the error of monitor task.
+            try:
+                fut.result()
+            except asyncio.CancelledError:
+                pass
+
+        # Create a monitor task to update progress and collect garbage.
+        monitor_task = asyncio.create_task(
+            self._update_progress_and_collect_garbage(
+                subtask_graph, self._config.get_subtask_monitor_interval()
+            )
+        )
+        monitor_task.add_done_callback(_on_monitor_task_done)
+
+        def _on_execute_task_done(fut):
+            # Make sure the monitor task is cancelled.
+            monitor_task.cancel()
+            # Just use `self._cur_stage_tile_progress` as current stage progress
+            # because current stage is completed, its progress is 1.0.
+            self._cur_stage_progress = 1.0
+            self._pre_all_stages_progress += self._cur_stage_tile_progress
+            self._cur_stage_first_output_object_ref_to_subtask.clear()
+
         self._execute_subtask_graph_aiotask = asyncio.current_task()
+        self._execute_subtask_graph_aiotask.add_done_callback(_on_execute_task_done)
 
         logger.info("Stage %s start.", stage_id)
         task_context = self._task_context
@@ -359,8 +381,9 @@ async def execute_subtask_graph(
                 continue
             elif output_count == 1:
                 output_object_refs = [output_object_refs]
-            self._cur_stage_output_object_refs.extend(output_object_refs)
-            self._output_object_refs_nums.append(len(output_object_refs))
+            self._cur_stage_first_output_object_ref_to_subtask[
+                output_object_refs[0]
+            ] = subtask
             if output_meta_keys:
                 meta_object_ref, *output_object_refs = output_object_refs
                 # TODO(fyrestone): Fetch(not get) meta object here.
@@ -395,16 +418,16 @@ async def execute_subtask_graph(
         logger.info("Waiting for stage %s complete.", stage_id)
         # Patched the asyncio.to_thread for Python < 3.9 at mars/lib/aio/__init__.py
         await asyncio.to_thread(ray.wait, list(output_object_refs), fetch_local=False)
-        # Just use `self._cur_stage_tile_progress` as current stage progress
-        # because current stage is finished, its progress is 1.
-        self._pre_all_stages_progress += self._cur_stage_tile_progress
-        self._cur_stage_output_object_refs.clear()
-        self._output_object_refs_nums.clear()
+
         logger.info("Stage %s is complete.", stage_id)
         return chunk_to_meta
 
     async def __aexit__(self, exc_type, exc_val, exc_tb):
         if exc_type is not None:
+            try:
+                await self.cancel()
+            except BaseException:  # noqa: E722  # nosec  # pylint: disable=bare-except
+                pass
             return
 
         # Update info if no exception occurs.
@@ -458,19 +481,7 @@ async def get_available_band_resources(self) -> Dict[BandType, Resource]:
 
     async def get_progress(self) -> float:
         """Get the execution progress."""
-        stage_progress = 0.0
-        total = len(self._cur_stage_output_object_refs)
-        if total > 0:
-            finished_objects, _ = ray.wait(
-                self._cur_stage_output_object_refs,
-                num_returns=total,
-                timeout=0,  # Avoid blocking the asyncio loop.
-                fetch_local=False,
-            )
-            stage_progress = (
-                len(finished_objects) / total * self._cur_stage_tile_progress
-            )
-        return self._pre_all_stages_progress + stage_progress
+        return self._cur_stage_progress
 
     async def cancel(self):
         """
@@ -480,26 +491,17 @@ async def cancel(self):
         2. Try to cancel the submitted subtasks by `ray.cancel`
         """
         logger.info("Start to cancel task %s.", self._task)
-        if self._task is None:
+        if self._task is None or self._cancelled is True:
             return
         self._cancelled = True
-        if (
-            self._execute_subtask_graph_aiotask is not None
-            and not self._execute_subtask_graph_aiotask.cancelled()
-        ):
+        if self._execute_subtask_graph_aiotask is not None:
             self._execute_subtask_graph_aiotask.cancel()
         timeout = self._config.get_subtask_cancel_timeout()
-        subtask_num = len(self._output_object_refs_nums)
-        if subtask_num > 0:
-            pos = 0
-            obj_refs_to_be_cancelled_ = []
-            for i in range(0, subtask_num):
-                if i > 0:
-                    pos += self._output_object_refs_nums[i - 1]
-                obj_refs_to_be_cancelled_.append(
-                    _cancel_ray_task(self._cur_stage_output_object_refs[pos], timeout)
-                )
-            await asyncio.gather(*obj_refs_to_be_cancelled_)
+        to_be_cancelled_coros = [
+            _cancel_ray_task(object_ref, timeout)
+            for object_ref in self._cur_stage_first_output_object_ref_to_subtask.keys()
+        ]
+        await asyncio.gather(*to_be_cancelled_coros)
 
     async def _load_subtask_inputs(
         self, stage_id: str, subtask: Subtask, chunk_graph: ChunkGraph, context: Dict
@@ -551,3 +553,81 @@ def _get_subtask_output_keys(chunk_graph: ChunkGraph):
             else:
                 output_keys[chunk.key] = 1
         return output_keys.keys()
+
+    async def _update_progress_and_collect_garbage(
+        self, subtask_graph: SubtaskGraph, interval_seconds: float
+    ):
+        object_ref_to_subtask = self._cur_stage_first_output_object_ref_to_subtask
+        total = len(subtask_graph)
+        completed_subtasks = OrderedSet()
+
+        def gc():
+            """
+            Consume the completed subtasks and collect garbage.
+
+            GC the output object refs of the subtask which successors are submitted
+            (not completed as above) can reduce the memory peaks, but we can't cancel
+            and rerun slow subtasks because the input object refs of running subtasks
+            may be deleted.
+            """
+            i = 0
+            gc_subtasks = set()
+
+            while i < total:
+                while i >= len(completed_subtasks):
+                    yield
+                # Iterate the completed subtasks once.
+                subtask = completed_subtasks[i]
+                i += 1
+                logger.debug("GC: %s", subtask)
+
+                # Note: There may be a scenario in which delayed gc occurs.
+                # When a subtask has more than one predecessor, like A, B,
+                # and in the `for ... in ...` loop we get A firstly while
+                # B's successors are completed, A's not. Then we cannot remove
+                # B's results chunks before A's.
+                for pred in subtask_graph.iter_predecessors(subtask):
+                    if pred in gc_subtasks:
+                        continue
+                    while not all(
+                        succ in completed_subtasks
+                        for succ in subtask_graph.iter_successors(pred)
+                    ):
+                        yield
+                    for chunk in pred.chunk_graph.results:
+                        self._task_context.pop(chunk.key, None)
+                    gc_subtasks.add(pred)
+
+            # TODO(fyrestone): Check the remaining self._task_context.keys()
+            # in the result subtasks
+
+        collect_garbage = gc()
+
+        while len(completed_subtasks) != total:
+            if len(object_ref_to_subtask) <= 0:  # pragma: no cover
+                await asyncio.sleep(interval_seconds)
+
+            # Only wait for unready subtask object refs.
+            ready_objects, _ = await asyncio.to_thread(
+                ray.wait,
+                list(object_ref_to_subtask.keys()),
+                num_returns=len(object_ref_to_subtask),
+                timeout=0,
+                fetch_local=False,
+            )
+            if len(ready_objects) == 0:
+                await asyncio.sleep(interval_seconds)
+                continue
+
+            # Pop the completed subtasks from object_ref_to_subtask.
+            completed_subtasks.update(map(object_ref_to_subtask.pop, ready_objects))
+            # Update progress.
+            stage_progress = (
+                len(completed_subtasks) / total * self._cur_stage_tile_progress
+            )
+            self._cur_stage_progress = self._pre_all_stages_progress + stage_progress
+            # Collect garbage, use `for ... in ...` to avoid raising StopIteration.
+            for _ in collect_garbage:
+                break
+            # Fast to next loop and give it a chance to update object_ref_to_subtask.
+            await asyncio.sleep(0)