""" Asynchronous Shared-Memory Scheduler for Dask Graphs. This scheduler coordinates several workers to execute tasks in a dask graph in parallel. It depends on a ``concurrent.futures.Executor`` and a corresponding Queue for worker-to-scheduler communication. It tries to execute tasks in an order which maintains a small memory footprint throughout execution. It does this by running tasks that allow us to release data resources. Task Selection Policy ===================== When we complete a task we add more data in to our set of available data; this new data makes new tasks available. We preferentially choose tasks that were just made available in a last-in-first-out fashion. We implement this as a simple stack. This results in more depth-first rather than breadth first behavior which encourages us to process batches of data to completion before starting in on new data when possible. When the addition of new data readies multiple tasks simultaneously we add tasks to the stack in sorted order so that tasks with greater keynames are run first. This can be handy to break ties in a predictable fashion. State ===== Many functions pass around a ``state`` variable that holds the current state of the computation. This variable consists of several other dictionaries and sets, explained below. Constant state -------------- 1. dependencies: {x: [a, b ,c]} a,b,c, must be run before x 2. dependents: {a: [x, y]} a must run before x or y Changing state -------------- ### Data 1. cache: available concrete data. {key: actual-data} 2. released: data that we've seen, used, and released because it is no longer needed ### Jobs 1. ready: A fifo stack of ready-to-run tasks 2. running: A set of tasks currently in execution 3. finished: A set of finished tasks 4. waiting: which tasks are still waiting on others :: {key: {keys}} Real-time equivalent of dependencies 5. waiting_data: available data to yet-to-be-run-tasks :: {key: {keys}} Real-time equivalent of dependents Examples -------- >>> import pprint # doctest: +SKIP >>> inc = lambda x: x + 1 >>> add = lambda x, y: x + y >>> dsk = {'x': 1, 'y': 2, 'z': (inc, 'x'), 'w': (add, 'z', 'y')} # doctest: +SKIP >>> pprint.pprint(start_state_from_dask(dsk)) # doctest: +SKIP {'cache': {'x': 1, 'y': 2}, 'dependencies': {'w': {'z', 'y'}, 'x': set(), 'y': set(), 'z': {'x'}}, 'dependents': defaultdict(None, {'w': set(), 'x': {'z'}, 'y': {'w'}, 'z': {'w'}}), 'finished': set(), 'ready': ['z'], 'released': set(), 'running': set(), 'waiting': {'w': {'z'}}, 'waiting_data': {'x': {'z'}, 'y': {'w'}, 'z': {'w'}}} Optimizations ============= We build this scheduler with out-of-core array operations in mind. To this end we have encoded some particular optimizations. Compute to release data ----------------------- When we choose a new task to execute we often have many options. Policies at this stage are cheap and can significantly impact performance. One could imagine policies that expose parallelism, drive towards a particular output, etc.. Our current policy is to run tasks that were most recently made available. Inlining computations --------------------- We hold on to intermediate computations either in memory or on disk. For very cheap computations that may emit new copies of the data, like ``np.transpose`` or possibly even ``x + 1`` we choose not to store these as separate pieces of data / tasks. Instead we combine them with the computations that require them. This may result in repeated computation but saves significantly on space and computation complexity. See the function ``inline_functions`` for more information. """ from __future__ import annotations import os from collections.abc import Mapping, Sequence from concurrent.futures import Executor, Future from functools import partial from queue import Empty, Queue from dask import config from dask.callbacks import local_callbacks, unpack_callbacks from dask.core import _execute_task, flatten, get_dependencies, has_tasks, reverse_dict from dask.order import order from dask.typing import Key if os.name == "nt": # Python 3 windows Queue.get doesn't handle interrupts properly. To # workaround this we poll at a sufficiently large interval that it # shouldn't affect performance, but small enough that users trying to kill # an application shouldn't care. def queue_get(q): while True: try: return q.get(block=True, timeout=0.1) except Empty: pass else: def queue_get(q): return q.get() def start_state_from_dask(dsk, cache=None, sortkey=None): """Start state from a dask Examples -------- >>> inc = lambda x: x + 1 >>> add = lambda x, y: x + y >>> dsk = {'x': 1, 'y': 2, 'z': (inc, 'x'), 'w': (add, 'z', 'y')} # doctest: +SKIP >>> from pprint import pprint # doctest: +SKIP >>> pprint(start_state_from_dask(dsk)) # doctest: +SKIP {'cache': {'x': 1, 'y': 2}, 'dependencies': {'w': {'z', 'y'}, 'x': set(), 'y': set(), 'z': {'x'}}, 'dependents': defaultdict(None, {'w': set(), 'x': {'z'}, 'y': {'w'}, 'z': {'w'}}), 'finished': set(), 'ready': ['z'], 'released': set(), 'running': set(), 'waiting': {'w': {'z'}}, 'waiting_data': {'x': {'z'}, 'y': {'w'}, 'z': {'w'}}} """ if sortkey is None: sortkey = order(dsk).get if cache is None: cache = config.get("cache", None) if cache is None: cache = dict() data_keys = set() for k, v in dsk.items(): if not has_tasks(dsk, v): cache[k] = v data_keys.add(k) dsk2 = dsk.copy() dsk2.update(cache) dependencies = {k: get_dependencies(dsk2, k) for k in dsk} waiting = {k: v.copy() for k, v in dependencies.items() if k not in data_keys} dependents = reverse_dict(dependencies) for a in cache: for b in dependents.get(a, ()): waiting[b].remove(a) waiting_data = {k: v.copy() for k, v in dependents.items() if v} ready_set = {k for k, v in waiting.items() if not v} ready = sorted(ready_set, key=sortkey, reverse=True) waiting = {k: v for k, v in waiting.items() if v} state = { "dependencies": dependencies, "dependents": dependents, "waiting": waiting, "waiting_data": waiting_data, "cache": cache, "ready": ready, "running": set(), "finished": set(), "released": set(), } return state """ Running tasks ------------- When we execute tasks we both 1. Perform the actual work of collecting the appropriate data and calling the function 2. Manage administrative state to coordinate with the scheduler """ def execute_task(key, task_info, dumps, loads, get_id, pack_exception): """ Compute task and handle all administration See Also -------- _execute_task : actually execute task """ try: task, data = loads(task_info) result = _execute_task(task, data) id = get_id() result = dumps((result, id)) failed = False except BaseException as e: result = pack_exception(e, dumps) failed = True return key, result, failed def batch_execute_tasks(it): """ Batch computing of multiple tasks with `execute_task` """ return [execute_task(*a) for a in it] def release_data(key, state, delete=True): """Remove data from temporary storage See Also -------- finish_task """ if key in state["waiting_data"]: assert not state["waiting_data"][key] del state["waiting_data"][key] state["released"].add(key) if delete: del state["cache"][key] def finish_task( dsk, key, state, results, sortkey, delete=True, release_data=release_data ): """ Update execution state after a task finishes Mutates. This should run atomically (with a lock). """ for dep in sorted(state["dependents"][key], key=sortkey, reverse=True): s = state["waiting"][dep] s.remove(key) if not s: del state["waiting"][dep] state["ready"].append(dep) for dep in state["dependencies"][key]: if dep in state["waiting_data"]: s = state["waiting_data"][dep] s.remove(key) if not s and dep not in results: release_data(dep, state, delete=delete) elif delete and dep not in results: release_data(dep, state, delete=delete) state["finished"].add(key) state["running"].remove(key) return state def nested_get(ind, coll): """Get nested index from collection Examples -------- >>> nested_get(1, 'abc') 'b' >>> nested_get([1, 0], 'abc') ('b', 'a') >>> nested_get([[1, 0], [0, 1]], 'abc') (('b', 'a'), ('a', 'b')) """ if isinstance(ind, list): return tuple(nested_get(i, coll) for i in ind) else: return coll[ind] def default_get_id(): """Default get_id""" return None def default_pack_exception(e, dumps): raise def reraise(exc, tb=None): if exc.__traceback__ is not tb: raise exc.with_traceback(tb) raise exc def identity(x): """Identity function. Returns x. >>> identity(3) 3 """ return x """ Task Selection -------------- We often have a choice among many tasks to run next. This choice is both cheap and can significantly impact performance. We currently select tasks that have recently been made ready. We hope that this first-in-first-out policy reduces memory footprint """ """ `get` ----- The main function of the scheduler. Get is the main entry point. """ def get_async( submit, num_workers, dsk, result, cache=None, get_id=default_get_id, rerun_exceptions_locally=None, pack_exception=default_pack_exception, raise_exception=reraise, callbacks=None, dumps=identity, loads=identity, chunksize=None, **kwargs, ): """Asynchronous get function This is a general version of various asynchronous schedulers for dask. It takes a ``concurrent.futures.Executor.submit`` function to form a more specific ``get`` method that walks through the dask array with parallel workers, avoiding repeat computation and minimizing memory use. Parameters ---------- submit : function A ``concurrent.futures.Executor.submit`` function num_workers : int The number of workers that task submissions can be spread over dsk : dict A dask dictionary specifying a workflow result : key or list of keys Keys corresponding to desired data cache : dict-like, optional Temporary storage of results get_id : callable, optional Function to return the worker id, takes no arguments. Examples are `threading.current_thread` and `multiprocessing.current_process`. rerun_exceptions_locally : bool, optional Whether to rerun failing tasks in local process to enable debugging (False by default) pack_exception : callable, optional Function to take an exception and ``dumps`` method, and return a serialized tuple of ``(exception, traceback)`` to send back to the scheduler. Default is to just raise the exception. raise_exception : callable, optional Function that takes an exception and a traceback, and raises an error. callbacks : tuple or list of tuples, optional Callbacks are passed in as tuples of length 5. Multiple sets of callbacks may be passed in as a list of tuples. For more information, see the dask.diagnostics documentation. dumps: callable, optional Function to serialize task data and results to communicate between worker and parent. Defaults to identity. loads: callable, optional Inverse function of `dumps`. Defaults to identity. chunksize: int, optional Size of chunks to use when dispatching work. Defaults to 1. If -1, will be computed to evenly divide ready work across workers. See Also -------- threaded.get """ chunksize = chunksize or config.get("chunksize", 1) queue = Queue() if isinstance(result, list): result_flat = set(flatten(result)) else: result_flat = {result} results = set(result_flat) dsk = dict(dsk) with local_callbacks(callbacks) as callbacks: _, _, pretask_cbs, posttask_cbs, _ = unpack_callbacks(callbacks) started_cbs = [] succeeded = False # if start_state_from_dask fails, we will have something # to pass to the final block. state = {} try: for cb in callbacks: if cb[0]: cb[0](dsk) started_cbs.append(cb) keyorder = order(dsk) state = start_state_from_dask(dsk, cache=cache, sortkey=keyorder.get) for _, start_state, _, _, _ in callbacks: if start_state: start_state(dsk, state) if rerun_exceptions_locally is None: rerun_exceptions_locally = config.get("rerun_exceptions_locally", False) if state["waiting"] and not state["ready"]: raise ValueError("Found no accessible jobs in dask") def fire_tasks(chunksize): """Fire off a task to the thread pool""" # Determine chunksize and/or number of tasks to submit nready = len(state["ready"]) if chunksize == -1: ntasks = nready chunksize = -(ntasks // -num_workers) else: used_workers = -(len(state["running"]) // -chunksize) avail_workers = max(num_workers - used_workers, 0) ntasks = min(nready, chunksize * avail_workers) # Prep all ready tasks for submission args = [] for _ in range(ntasks): # Get the next task to compute (most recently added) key = state["ready"].pop() # Notify task is running state["running"].add(key) for f in pretask_cbs: f(key, dsk, state) # Prep args to send data = { dep: state["cache"][dep] for dep in get_dependencies(dsk, key) } args.append( ( key, dumps((dsk[key], data)), dumps, loads, get_id, pack_exception, ) ) # Batch submit for i in range(-(len(args) // -chunksize)): each_args = args[i * chunksize : (i + 1) * chunksize] if not each_args: break fut = submit(batch_execute_tasks, each_args) fut.add_done_callback(queue.put) # Main loop, wait on tasks to finish, insert new ones while state["waiting"] or state["ready"] or state["running"]: fire_tasks(chunksize) for key, res_info, failed in queue_get(queue).result(): if failed: exc, tb = loads(res_info) if rerun_exceptions_locally: data = { dep: state["cache"][dep] for dep in get_dependencies(dsk, key) } task = dsk[key] _execute_task(task, data) # Re-execute locally else: raise_exception(exc, tb) res, worker_id = loads(res_info) state["cache"][key] = res finish_task(dsk, key, state, results, keyorder.get) for f in posttask_cbs: f(key, res, dsk, state, worker_id) succeeded = True finally: for _, _, _, _, finish in started_cbs: if finish: finish(dsk, state, not succeeded) return nested_get(result, state["cache"]) """ Synchronous concrete version of get_async Usually we supply a ``concurrent.futures.Executor``. Here we provide a sequential one. This is useful for debugging and for code dominated by the GIL """ class SynchronousExecutor(Executor): _max_workers = 1 def submit(self, fn, *args, **kwargs): fut = Future() try: fut.set_result(fn(*args, **kwargs)) except BaseException as e: fut.set_exception(e) return fut synchronous_executor = SynchronousExecutor() def get_sync(dsk: Mapping, keys: Sequence[Key] | Key, **kwargs): """A naive synchronous version of get_async Can be useful for debugging. """ kwargs.pop("num_workers", None) # if num_workers present, remove it return get_async( synchronous_executor.submit, synchronous_executor._max_workers, dsk, keys, **kwargs, ) """ Adaptor for ``multiprocessing.Pool`` instances Usually we supply a ``concurrent.futures.Executor``. Here we provide a wrapper class for ``multiprocessing.Pool`` instances so we can treat them like ``concurrent.futures.Executor`` instances instead. This is mainly useful for legacy use cases or users that prefer ``multiprocessing.Pool``. """ class MultiprocessingPoolExecutor(Executor): def __init__(self, pool): self.pool = pool self._max_workers = len(pool._pool) def submit(self, fn, *args, **kwargs): return submit_apply_async(self.pool.apply_async, fn, *args, **kwargs) def submit_apply_async(apply_async, fn, *args, **kwargs): fut = Future() apply_async(fn, args, kwargs, fut.set_result, fut.set_exception) return fut def get_apply_async(apply_async, num_workers, *args, **kwargs): return get_async( partial(submit_apply_async, apply_async), num_workers, *args, **kwargs ) def sortkey(item): """Sorting key function that is robust to different types Both strings and tuples are common key types in dask graphs. However In Python 3 one can not compare strings with tuples directly. This function maps many types to a form where they can be compared Examples -------- >>> sortkey('Hello') ('str', 'Hello') >>> sortkey(('x', 1)) ('tuple', ('x', 1)) """ return (type(item).__name__, item)