Active Memory Manager framework + discard excess replicas

distributed/active_memory_manager.py

@@ -0,0 +1,305 @@
+from __future__ import annotations
+
+import asyncio
+from collections import defaultdict
+from collections.abc import Generator
+from typing import TYPE_CHECKING, Optional
+
+from tornado.ioloop import PeriodicCallback
+
+import dask
+from dask.utils import parse_timedelta
+
+from .utils import import_term
+
+if TYPE_CHECKING:
+    from .scheduler import SchedulerState, TaskState, WorkerState
+
+
+class ActiveMemoryManagerExtension:
+    """Scheduler extension that optimizes memory usage across the cluster.
+    It can be either triggered by hand or automatically every few seconds; at every
+    iteration it performs one or both of the following:
+
+    - create new replicas of in-memory tasks
+    - destroy replicas of in-memory tasks; this never destroys the last available copy.
+
+    There are no 'move' operations. A move is performed in two passes: first you create
+    a copy and, in the next iteration, you delete the original (if the copy succeeded).
+
+    This extension is configured by the dask config section
+    ``distributed.scheduler.active-memory-manager``.
+    """
+
+    scheduler: SchedulerState
+    policies: set[ActiveMemoryManagerPolicy]
+    interval: float
+
+    # These attributes only exist within the scope of self.run()
+    # Current memory (in bytes) allocated on each worker, plus/minus pending actions
+    workers_memory: dict[WorkerState, int]
+    # Pending replications and deletions for each task
+    pending: defaultdict[TaskState, tuple[set[WorkerState], set[WorkerState]]]
+
+    def __init__(
+        self,
+        scheduler: SchedulerState,
+        # The following parameters are exposed so that one may create, run, and throw
+        # away on the fly a specialized manager, separate from the main one.
+        policies: set[ActiveMemoryManagerPolicy] = None,
+        register: bool = True,
+        start: bool = None,
+        interval: float = None,
+    ):
+        self.scheduler = scheduler
+
+        if policies is None:
+            policies = set()
+            for kwargs in dask.config.get(
+                "distributed.scheduler.active-memory-manager.policies"
+            ):
+                kwargs = kwargs.copy()
+                cls = import_term(kwargs.pop("class"))
+                if not issubclass(cls, ActiveMemoryManagerPolicy):
+                    raise TypeError(
+                        f"{cls}: Expected ActiveMemoryManagerPolicy; got {type(cls)}"
+                    )
+                policies.add(cls(**kwargs))
+
+        for policy in policies:
+            policy.manager = self
+        self.policies = policies
+
+        if register:
+            scheduler.extensions["amm"] = self
+            scheduler.handlers.update(
+                {
+                    "amm_run_once": self.run_once,
+                    "amm_start": self.start,
+                    "amm_stop": self.stop,
+                }
+            )
+
+        if interval is None:
+            interval = parse_timedelta(
+                dask.config.get("distributed.scheduler.active-memory-manager.interval")
+            )
+        self.interval = interval
+        if start is None:
+            start = dask.config.get("distributed.scheduler.active-memory-manager.start")
+        if start:
+            self.start()
+
+    def start(self, comm=None) -> None:
+        """Start executing very <interval> seconds until scheduler shutdown"""
+        pc = PeriodicCallback(self.run_once, self.interval * 1000.0)
+        self.scheduler.periodic_callbacks["amm"] = pc
+        pc.start()
+
+    def stop(self, comm=None) -> None:
+        """Stop periodic execution"""
+        pc = self.scheduler.periodic_callbacks.pop("amm", None)
+        if pc:
+            pc.stop()
+
+    def run_once(self, comm=None) -> None:
+        """Run all policies once and asynchronously (fire and forget) enact their
+        recommendations to replicate/drop keys
+        """
+        # This should never fail since this is a synchronous method
+        assert not hasattr(self, "pending")
+
+        self.pending = defaultdict(lambda: (set(), set()))
+        self.workers_memory = {
+            w: w.memory.optimistic for w in self.scheduler.workers.values()
+        }
+        try:
+            # populate self.pending
+            self._run_policies()
+
+            drop_by_worker = defaultdict(set)
+            repl_by_worker = defaultdict(dict)
+            for ts, (pending_repl, pending_drop) in self.pending.items():
+                if not ts.who_has:
+                    continue
+                who_has = [ws_snd.address for ws_snd in ts.who_has - pending_drop]
+                assert who_has  # Never drop the last replica
+                for ws_rec in pending_repl:
+                    assert ws_rec not in ts.who_has
+                    repl_by_worker[ws_rec.address][ts.key] = who_has
+                for ws in pending_drop:
+                    assert ws in ts.who_has
+                    drop_by_worker[ws.address].add(ts.key)
+
+            # Fire-and-forget enact recommendations from policies
+            # This is temporary code, waiting for
+            # https://github.com/dask/distributed/pull/5046
+            for addr, who_has in repl_by_worker.items():
+                asyncio.create_task(self.scheduler.gather_on_worker(addr, who_has))
+            for addr, keys in drop_by_worker.items():
+                asyncio.create_task(self.scheduler.delete_worker_data(addr, keys))
+            # End temporary code
+
+        finally:
+            del self.workers_memory
+            del self.pending
+
+    def _run_policies(self) -> None:
+        """Sequentially run ActiveMemoryManagerPolicy.run() for all registered policies,
+        obtain replicate/drop suggestions, and use them to populate self.pending.
+        """
+        candidates: Optional[set[WorkerState]]
+        cmd: str
+        ws: Optional[WorkerState]
+        ts: TaskState
+        nreplicas: int
+
+        for policy in list(self.policies):  # a policy may remove itself
+            policy_gen = policy.run()
+            ws = None
+            while True:
+                try:
+                    cmd, ts, candidates = policy_gen.send(ws)
+                except StopIteration:
+                    break  # next policy
+
+                pending_repl, pending_drop = self.pending[ts]
+
+                if cmd == "replicate":
+                    ws = self._find_recipient(ts, candidates, pending_repl)
+                    if ws:
+                        pending_repl.add(ws)
+                        self.workers_memory[ws] += ts.nbytes
+
+                elif cmd == "drop":
+                    ws = self._find_dropper(ts, candidates, pending_drop)
+                    if ws:
+                        pending_drop.add(ws)
+                        self.workers_memory[ws] = max(
+                            0, self.workers_memory[ws] - ts.nbytes
+                        )
+
+                else:
+                    raise ValueError(f"Unknown command: {cmd}")  # pragma: nocover
+
+    def _find_recipient(
+        self,
+        ts: TaskState,
+        candidates: Optional[set[WorkerState]],
+        pending_repl: set[WorkerState],
+    ) -> Optional[WorkerState]:
+        """Choose a worker to acquire a new replica of an-in-memory task among a set of
+        candidates. If candidates is None, default to all workers in the cluster that do
+        not hold a replica yet. The worker with the lowest memory usage (downstream of
+        pending replications and drops) will be returned.
+        """
+        if ts.state != "memory":
+            return None
+        if candidates is None:
+            candidates = set(self.scheduler.workers.values())
+        candidates -= ts.who_has
+        candidates -= pending_repl
+        if not candidates:
+            return None
+        # id(ws) is there just to prevent WorkerState objects to be compared in the
+        # unlikely event that they have exactly the same amount of bytes allocated.
+        _, _, ws = min((self.workers_memory[ws], id(ws), ws) for ws in candidates)
+        return ws
+
+    def _find_dropper(
+        self,
+        ts: TaskState,
+        candidates: Optional[set[WorkerState]],
+        pending_drop: set[WorkerState],
+    ) -> Optional[WorkerState]:
+        """Choose a worker to drop its replica of an in-memory task among a set of
+        candidates. If candidates is None, default to all workers in the cluster that
+        hold a replica. The worker with the highest memory usage (downstream of pending
+        replications and drops) will be returned.
+        """
+        if len(ts.who_has) - len(pending_drop) < 2:
+            return None
+        if candidates is None:
+            candidates = ts.who_has.copy()
+        else:
+            candidates &= ts.who_has
+        candidates -= pending_drop
+        candidates -= {waiter_ts.processing_on for waiter_ts in ts.waiters}
+        if not candidates:
+            return None
+        # id(ws) is there just to prevent WorkerState objects to be compared in the
+        # unlikely event that they have exactly the same amount of bytes allocated.
+        _, _, ws = max((self.workers_memory[ws], id(ws), ws) for ws in candidates)
+        return ws
+
+
+class ActiveMemoryManagerPolicy:
+    """Abstract parent class"""
+
+    manager: ActiveMemoryManagerExtension
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}()"
+
+    def run(
+        self,
+    ) -> Generator[
+        tuple[str, TaskState, Optional[set[WorkerState]]],
+        Optional[WorkerState],
+        None,
+    ]:
+        """This method is invoked by the ActiveMemoryManager every few seconds, or
+        whenever the user invokes scheduler.amm_run_once().
+        It is an iterator that must emit any of the following:
+
+        - "replicate", <TaskState>, None
+        - "replicate", <TaskState>, {subset of potential workers to replicate to}
+        - "drop", <TaskState>, None
+        - "drop", <TaskState>, {subset of potential workers to drop from}
+
+        Each element yielded indicates the desire to create or destroy a single replica
+        of a key. If a subset of workers is not provided, it defaults to all workers on
+        the cluster. Either the ActiveMemoryManager or the Worker may later decide to
+        disregard the request, e.g. because it would delete the last copy of a key or
+        because the key is currently needed on that worker.
+
+        You may optionally retrieve which worker it was decided the key will be
+        replicated to or dropped from, as follows:
+
+        ```python
+        choice = yield "replicate", ts, None
+        ```
+
+        ``choice`` is either a WorkerState or None; the latter is returned if the
+        ActiveMemoryManager chose to disregard the request.
+
+        The current pending (accepted) commands can be inspected on
+        ``self.manager.pending``; this includes the commands previously yielded by this
+        same method.
+
+        The current memory usage on each worker, *downstream of all pending commands*,
+        can be inspected on ``self.manager.workers_memory``.
+        """
+        raise NotImplementedError("Virtual method")
+
+
+class ReduceReplicas(ActiveMemoryManagerPolicy):
+    """Make sure that in-memory tasks are not replicated on more workers than desired;
+    drop the excess replicas.
+    """
+
+    def run(self):
+        # TODO this is O(n) to the total number of in-memory tasks on the cluster; it
+        #      could be made faster by automatically attaching it to a TaskState when it
+        #      goes above one replica and detaching it when it drops below two.
+        for ts in self.manager.scheduler.tasks.values():
+            if len(ts.who_has) < 2:
+                continue
+
+            desired_replicas = 1  # TODO have a marker on TaskState
+            ndrop = len(ts.who_has) - max(desired_replicas, len(ts.waiters))
+            if ts in self.manager.pending:
+                pending_repl, pending_drop = self.manager.pending[ts]
+                ndrop += len(pending_repl) - len(pending_drop)
+            for _ in range(ndrop):
+                yield "drop", ts, None

distributed/distributed-schema.yaml

@@ -242,6 +242,31 @@ properties:
               A list of trusted root modules the schedular is allowed to import (incl. submodules). For security reasons, the
               scheduler does not import arbitrary Python modules.
 
+          active-memory-manager:
+            type: object
+            required: [start, interval, policies]
+            additionalProperties: false
+            properties:
+              start:
+                type: boolean
+                description: set to true to auto-start the AMM on Scheduler init;
+                  false to manually start it with client.scheduler.amm_start()
+              interval:
+                type: string
+                description:
+                  Time expression, e.g. "2s". Run the AMM cycle every <interval>.
+              policies:
+                type: array
+                items:
+                  type: object
+                  required: [class]
+                  properties:
+                    class:
+                      type: string
+                      description: fully qualified name of an ActiveMemoryManagerPolicy
+                        subclass
+                  additionalProperties:
+                    description: keyword arguments to the policy constructor, if any
 
       worker:
         type: object

distributed/distributed.yaml

@@ -60,6 +60,18 @@ distributed:
       - dask
       - distributed
 
+    active-memory-manager:
+      # Set to true to auto-start the Active Memory Manager on Scheduler start; if false
+      # you'll have to either manually start it with client.scheduler.amm_start() or run
+      # it once with client.scheduler.amm_run().
+      start: false
+      # Once started, run the AMM cycle every <interval>
+      interval: 2s
+      policies:
+        # Policies that should be executed at every cycle. Any additional keys in each
+        # object are passed as keyword arguments to the policy constructor.
+        - class: distributed.active_memory_manager.ReduceReplicas
+
   worker:
     blocked-handlers: []
     multiprocessing-method: spawn