#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# pyre-strict
import logging
import random
import time
from dataclasses import dataclass
from logging import Logger
from ax.core.base_trial import TrialStatus
from ax.utils.common.base import Base
from ax.utils.common.logger import get_logger
from pyre_extensions import none_throws
logger: Logger = get_logger(__name__)
[docs]
@dataclass
class SimTrial:
"""Container for the simulation tasks.
Attributes:
trial_index: The index of the trial (should match Ax trial index).
sim_runtime: The runtime of the trial (sampled at creation).
sim_start_time: When the trial started running (or exits queued state).
sim_queued_time: When the trial was initially queued.
sim_completed_time: When the trial was marked as completed. Currently,
this is used by an early-stopper via ``stop_trial``.
"""
# The (Ax) trial index
trial_index: int
# The simulation runtime in seconds
sim_runtime: float
# the start time in seconds
sim_start_time: float | None = None
# the queued time in seconds
sim_queued_time: float | None = None
# the completed time (used for early stopping)
sim_completed_time: float | None = None
[docs]
@dataclass
class SimStatus:
"""Container for status of the simulation.
Attributes:
queued: List of indices of queued trials.
running: List of indices of running trials.
failed: List of indices of failed trials.
time_remaining: List of sim time remaining for running trials.
completed: List of indicies of completed trials.
"""
queued: list[int]
running: list[int]
failed: list[int]
time_remaining: list[float]
completed: list[int]
[docs]
@dataclass
class BackendSimulatorOptions:
"""Settings for the BackendSimulator.
Args:
max_concurrency: The maximum number of trials that can be run
in parallel.
time_scaling: The factor to scale down the runtime of the tasks by.
If ``runtime`` is the actual runtime of a trial, the simulation
time will be ``runtime / time_scaling``.
failure_rate: The rate at which the trials are failing. For now, trials
fail independently with at coin flip based on that rate.
internal_clock: The initial state of the internal clock. If `None`,
the simulator uses ``time.time()`` as the clock.
use_update_as_start_time: Whether the start time of a new trial should be logged
as the current time (at time of update) or end time of previous trial.
This makes sense when using the internal clock and the BackendSimulator
is simulated forward by an external process (such as Scheduler).
"""
max_concurrency: int = 1
time_scaling: float = 1.0
failure_rate: float = 0.0
internal_clock: float | None = None
use_update_as_start_time: bool = False
[docs]
@dataclass
class BackendSimulatorState:
"""State of the BackendSimulator.
Args:
options: The BackendSimulatorOptions associated with this simulator.
verbose_logging: Whether the simulator is using verbose logging.
queued: Currently queued trials.
running: Currently running trials.
failed: Currently failed trials.
completed: Currently completed trials.
"""
options: BackendSimulatorOptions
verbose_logging: bool
queued: list[dict[str, float | None]]
running: list[dict[str, float | None]]
failed: list[dict[str, float | None]]
completed: list[dict[str, float | None]]
[docs]
class BackendSimulator(Base):
"""Simulator for a backend deployment with concurrent dispatch and a queue."""
def __init__(
self,
options: BackendSimulatorOptions | None = None,
queued: list[SimTrial] | None = None,
running: list[SimTrial] | None = None,
failed: list[SimTrial] | None = None,
completed: list[SimTrial] | None = None,
verbose_logging: bool = True,
) -> None:
"""A simulator for a concurrent dispatch with a queue.
Args:
options: A ``BackendSimulatorOptions`` object with various settings
of the backend simulator.
queued: A list of SimTrial objects representing the queued trials
(only used for testing particular initialization cases)
running: A list of SimTrial objects representing the running trials
(only used for testing particular initialization cases)
failed: A list of SimTrial objects representing the failed trials
(only used for testing particular initialization cases)
completed: A list of SimTrial objects representing the completed trials
(only used for testing particular initialization cases)
verbose_logging: If False, sets the logging level to WARNING.
"""
if not verbose_logging:
logger.setLevel(logging.WARNING)
self.options: BackendSimulatorOptions = (
BackendSimulatorOptions() if options is None else options
)
self._queued: list[SimTrial] = queued or []
self._running: list[SimTrial] = running or []
self._failed: list[SimTrial] = failed or []
self._completed: list[SimTrial] = completed or []
self._verbose_logging = verbose_logging
self._create_index_to_trial_map()
@property
def num_queued(self) -> int:
"""The number of queued trials (to run as soon as capacity is available)."""
return len(self._queued)
@property
def num_running(self) -> int:
"""The number of currently running trials."""
return len(self._running)
@property
def num_failed(self) -> int:
"""The number of failed trials."""
return len(self._failed)
@property
def num_completed(self) -> int:
"""The number of completed trials."""
return len(self._completed)
@property
def max_concurrency(self) -> int:
"""The maximum number of trials that can be run in parallel."""
return self.options.max_concurrency
@property
def time_scaling(self) -> float:
"""The factor to scale down the runtime of the tasks by."""
return self.options.time_scaling
@property
def failure_rate(self) -> float:
"""The rate at which the trials randomly fail."""
return self.options.failure_rate
@property
def _internal_clock(self) -> float | None:
"""The internal clock of the simulator."""
return self.options.internal_clock
@property
def use_internal_clock(self) -> bool:
"""Whether or not we are using the internal clock."""
return self._internal_clock is not None
@property
def time(self) -> float:
"""The current time."""
if self.use_internal_clock:
return none_throws(self._internal_clock)
return time.time()
@property
def all_trials(self) -> list[SimTrial]:
"""All trials on the simulator."""
return self._queued + self._running + self._completed + self._failed
[docs]
def update(self) -> None:
"""Update the state of the simulator."""
if self.use_internal_clock:
self.options.internal_clock = none_throws(self.options.internal_clock) + 1
self._update(self.time)
state = self.state()
logger.info(
"\n-----------\n"
f"Updated backend simulator state (time = {self.time}):\n"
f"** Queued:\n{format(state.queued)}\n"
f"** Running:\n{format(state.running)}\n"
f"** Failed:\n{format(state.failed)}\n"
f"** Completed:\n{format(state.completed)}\n"
f"-----------\n"
)
[docs]
def state(self) -> BackendSimulatorState:
"""Return a ``BackendSimulatorState`` containing the state of the simulator."""
return BackendSimulatorState(
options=self.options,
verbose_logging=self._verbose_logging,
queued=[q.__dict__.copy() for q in self._queued],
running=[r.__dict__.copy() for r in self._running],
failed=[r.__dict__.copy() for r in self._failed],
completed=[c.__dict__.copy() for c in self._completed],
)
[docs]
def run_trial(self, trial_index: int, runtime: float) -> None:
"""Run a simulated trial.
Args:
trial_index: The index of the trial (usually the Ax trial index)
runtime: The runtime of the simulation. Typically sampled from the
runtime model of a simulation model.
Internally, the runtime is scaled by the `time_scaling` factor, so that
the simulation can run arbitrarily faster than the underlying evaluation.
"""
# scale runtime to simulation
sim_runtime = runtime / self.time_scaling
# flip a coin to see if the trial fails (for now fail instantly)
if self.failure_rate > 0:
if random.random() < self.failure_rate:
self._failed.append(
SimTrial(
trial_index=trial_index,
sim_runtime=sim_runtime,
sim_start_time=self.time,
)
)
return
if self.num_running < self.max_concurrency:
# note that though these are running for simulation purposes,
# the trial status does not yet get updated (this is also how it
# works in the real world, this requires updating the trial status manually)
curr_time = self.time
new_trial = SimTrial(
trial_index=trial_index,
sim_runtime=sim_runtime,
sim_start_time=curr_time,
sim_queued_time=curr_time,
)
self.new_trial(trial=new_trial, status=TrialStatus.RUNNING)
else:
new_trial = SimTrial(
trial_index=trial_index,
sim_runtime=sim_runtime,
sim_queued_time=self.time,
)
self.new_trial(trial=new_trial, status=TrialStatus.STAGED)
[docs]
def new_trial(self, trial: SimTrial, status: TrialStatus) -> None:
"""Register a trial into the simulator.
Args:
trial: A new trial to add.
status: The status of the new trial, either STAGED (add to ``self._queued``)
or RUNNING (add to ``self._running``).
"""
if status == TrialStatus.STAGED:
self._queued.append(trial)
elif status == TrialStatus.RUNNING:
self._running.append(trial)
else:
raise ValueError("New trials must be either staged or running.")
self._index_to_trial_map[trial.trial_index] = trial
[docs]
def stop_trial(self, trial_index: int) -> None:
"""Stop a simulated trial by setting the completed time to the current time.
Args:
trial_index: The index of the trial to stop.
"""
trial_status = self.lookup_trial_index_status(trial_index)
if trial_status is not TrialStatus.RUNNING:
logger.info(
f"Trial {trial_index} is not currently running (has status "
f"{trial_status}) and cannot be stopped."
)
else:
trial = self._index_to_trial_map[trial_index]
trial.sim_completed_time = self.time
logger.info(
f"Trial {trial_index} stopped at time {trial.sim_completed_time}."
)
[docs]
def status(self) -> SimStatus:
"""Return the internal status of the simulator.
Returns:
A ``SimStatus`` object representing the current simulator state.
"""
now = self.time
return SimStatus(
queued=[t.trial_index for t in self._queued],
running=[t.trial_index for t in self._running],
failed=[t.trial_index for t in self._failed],
time_remaining=[
# pyre-fixme[58]: `+` is not supported for operand types
# `Optional[float]` and `float`.
t.sim_start_time + t.sim_runtime - now
for t in self._running
],
completed=[t.trial_index for t in self._completed],
)
[docs]
def lookup_trial_index_status(self, trial_index: int) -> TrialStatus:
"""Lookup the trial status of a ``trial_index``.
Args:
trial_index: The index of the trial to check.
Returns:
A ``TrialStatus``.
"""
sim_status = self.status()
if trial_index in sim_status.queued:
return TrialStatus.STAGED
elif trial_index in sim_status.running:
return TrialStatus.RUNNING
elif trial_index in sim_status.completed:
return TrialStatus.COMPLETED
elif trial_index in sim_status.failed:
return TrialStatus.FAILED
raise ValueError(f"Trial {trial_index} not found in simulator.")
[docs]
def get_sim_trial_by_index(self, trial_index: int) -> SimTrial | None:
"""Get a ``SimTrial`` by ``trial_index``.
Args:
trial_index: The index of the trial to return.
Returns:
A ``SimTrial`` with the index ``trial_index`` or None if not found.
"""
return self._index_to_trial_map.get(trial_index)
def _update_completed(self, timestamp: float) -> list[SimTrial]:
"""Look through running trials and see if any trials have completed
since the last check. Such trials could have completed naturally (in
this case, ``sim_completed_time`` is None) or have been given a artificial
completion time (``sim_completed_time`` is not None) via early stopping.
Args:
timestamp: The current timestamp.
"""
completed_since_last = []
new_running = []
for trial in self._running:
# pyre-fixme[58]: `+` is not supported for operand types
# `Optional[float]` and `float`.
if timestamp >= trial.sim_start_time + trial.sim_runtime:
completed_since_last.append(trial)
trial.sim_completed_time = (
trial.sim_start_time + trial.sim_runtime # pyre-ignore[58]
)
elif (
trial.sim_completed_time is not None
and timestamp >= trial.sim_completed_time
):
completed_since_last.append(trial) # was early stopped
else:
new_running.append(trial)
self._running = new_running
self._completed.extend(completed_since_last)
return completed_since_last
def _update(self, timestamp: float) -> None:
"""Check if trials have completed (or stopped) and update the simulator.
Args:
timestamp: The current timestamp.
"""
completed_since_last = self._update_completed(timestamp)
# if no trial has finished since the last call we're done
if len(completed_since_last) == 0:
return
# if at least one trial has finished, we need to graduate queued trials to
# running trials. Since all we need to keep track of is the start_time, we can
# do this retroactively.
# TODO: Improve performance / make less ad hoc by using a priority queue
for c in completed_since_last:
if self.num_queued > 0:
new_running_trial = self._queued.pop(0)
sim_start_time = (
# pyre-fixme[58]: `+` is not supported for operand types
# `Optional[float]` and `float`.
c.sim_start_time + c.sim_runtime
if not self.options.use_update_as_start_time
else self.time
)
new_running_trial.sim_start_time = sim_start_time
self._running.append(new_running_trial)
# since these graduated trials could both have started and finished in between
# the simulation updates, we need to re-run the update with the new state
self._update(timestamp)
def _create_index_to_trial_map(self) -> None:
"""Create the index to trial map, which is useful for getting
the ``SimTrial`` objects, as in ``get_sim_trial_by_index``."""
self._index_to_trial_map = {t.trial_index: t for t in self.all_trials}
