1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258
|
"""Run benchmarks while handling parallelism, isolation, and fault tolerance."""
import math
import multiprocessing
import subprocess
import textwrap
import threading
import time
from typing import Dict, List, Optional, Set, Tuple, Union
from execution.work import PYTHON_CMD, SHELL, InProgress, WorkOrder
from worker.main import WorkerFailure, WorkerOutput
CPU_COUNT: int = multiprocessing.cpu_count()
class WorkerFailed(Exception):
"""Raised in the main process when a worker failure is detected."""
def __init__(self, cmd: str, wrapped_trace: Optional[str] = None) -> None:
self.cmd: str = cmd
self.wrapped_trace: Optional[str] = wrapped_trace
super().__init__()
class CorePool:
"""Allocator style helper class to assign individual tasks to a core range.
Pinning tasks to separate cores (or core ranges if `num_threads` > 1)
serves two purposes. First, it prevents the machine from being overloaded,
which can result in OOMs or Callgrind crashes. Second, it helps reduce
noise in the wall times, which are collected as a secondary metric. For
multi-threaded workloads, adjacency is important. Often pairs of cores
share silicon (e.g. cache), while far away cores may lie on separate NUMA
nodes. For this reason, CorePool will only allocate contiguous core ranges.
This falls short of full architecture awareness, and instead tries to find
a balance between rigor and engineering complexity.
"""
def __init__(self, min_core_id: int, max_core_id: int) -> None:
assert min_core_id >= 0
assert max_core_id >= min_core_id
assert max_core_id < CPU_COUNT
self._min_core_id: int = min_core_id
self._max_core_id: int = max_core_id
self._num_cores = max_core_id - min_core_id + 1
print(f"Core pool created: cores {self._min_core_id}-{self._max_core_id}")
self._available: List[bool] = [
True for _ in range(min_core_id, min_core_id + self._num_cores)]
self._reservations: Dict[str, Tuple[int, ...]] = {}
self._lock = threading.Lock()
def reserve(self, n: int) -> Optional[str]:
"""Simple first-fit policy.
If successful, return a string for `taskset`. Otherwise, return None.
"""
with self._lock:
for lower_index in range(self._num_cores - n + 1):
indices = tuple(range(lower_index, lower_index + n))
if all(self._available[i] for i in indices):
for i in indices:
self._available[i] = False
lower_core = indices[0] + self._min_core_id
upper_core = indices[-1] + self._min_core_id
key = f"{lower_core}-{upper_core}" if n > 1 else f"{lower_core}"
self._reservations[key] = indices
return key
return None
def release(self, key: str) -> None:
with self._lock:
for i in self._reservations[key]:
self._available[i] = True
self._reservations.pop(key)
class Runner:
def __init__(
self,
work_items: Tuple[WorkOrder, ...],
core_pool: Optional[CorePool] = None,
cadence: float = 1.0,
) -> None:
self._work_items: Tuple[WorkOrder, ...] = work_items
self._core_pool: CorePool = core_pool or CorePool(0, CPU_COUNT - 4)
self._cadence: float = cadence
# Working state.
self._work_queue: List[WorkOrder] = list(work_items)
self._active_jobs: List[InProgress] = []
self._results: Dict[WorkOrder, WorkerOutput] = {}
# Debug information for ETA and error messages.
self._start_time: float = -1
self._durations: Dict[WorkOrder, float] = {}
self._currently_processed: Optional[WorkOrder] = None
if len(work_items) != len(set(work_items)):
raise ValueError('Duplicate work items.')
def run(self) -> Dict[WorkOrder, WorkerOutput]:
try:
return self._run()
except KeyboardInterrupt:
print("\n\nKeyboardInterrupt (ctrl-c) detected. Shutting down children.")
self._force_shutdown(verbose=False)
raise
except subprocess.TimeoutExpired:
print("\n\nJob timed out. Shutting down children.")
self._force_shutdown(verbose=True)
raise
except WorkerFailed as e:
print('Shutting down all outstanding jobs before re-raising.')
self._force_shutdown(verbose=True)
print(f"Cmd: {e.cmd}")
if e.wrapped_trace:
print(e.wrapped_trace)
else:
print('Unknown failure. (Worker did not report exception contents.)')
raise
except BaseException:
print("\n\nUnknown exception. Shutting down jobs before re-raising.")
self._force_shutdown(verbose=True)
raise
def _run(self) -> Dict[WorkOrder, WorkerOutput]:
self._start_time = time.time()
self._canary_import()
while self._work_queue or self._active_jobs:
t0 = time.time()
self._update_active_jobs()
self._enqueue_new_jobs()
self._print_progress()
time.sleep(max(self._cadence - (time.time() - t0), 0.0))
print(f"\nTotal time: {time.time() - self._start_time:.0f} seconds")
return self._results.copy()
def _update_active_jobs(self) -> None:
active_jobs: List[InProgress] = []
for job in self._active_jobs:
self._currently_processed = job.work_order
if not job.check_finished():
active_jobs.append(job)
continue
result: Union[WorkerOutput, WorkerFailure] = job.result
if isinstance(result, WorkerOutput):
self._results[job.work_order] = result
assert job.cpu_list is not None
self._core_pool.release(job.cpu_list)
self._durations[job.work_order] = job.duration
else:
assert isinstance(result, WorkerFailure)
raise WorkerFailed(cmd=job.proc.cmd, wrapped_trace=result.failure_trace)
self._currently_processed = None
self._active_jobs.clear()
self._active_jobs.extend(active_jobs)
def _enqueue_new_jobs(self) -> None:
work_queue: List[WorkOrder] = []
for i, work_order in enumerate(self._work_queue):
self._currently_processed = work_order
cpu_list = self._core_pool.reserve(work_order.timer_args.num_threads)
if cpu_list is None:
work_queue.append(work_order)
else:
self._active_jobs.append(InProgress(work_order, cpu_list))
# Stagger creation. This helps with contention.
time.sleep(0.5)
self._currently_processed = None
self._work_queue.clear()
self._work_queue.extend(work_queue)
def _print_progress(self) -> None:
fraction = f"{len(self._results)} / {len(self._work_items)}"
elapsed = f"{time.time() - self._start_time:.0f} seconds"
if len(self._results) < 5:
eta = "Unknown"
else:
remaining = len(self._work_items) - len(self._results)
iters_remaining = math.ceil(remaining / self._core_pool._num_cores)
mean_time = sum(self._durations.values()) / len(self._durations)
eta_minutes = math.ceil(iters_remaining * mean_time / 60)
eta = f"~{eta_minutes:.0f} minute{'s' if eta_minutes > 1 else ''}"
print(f"\r{fraction} ({elapsed}), ETA: {eta}", end="")
def _force_shutdown(self, verbose: bool = False) -> None:
"""Try to interrupt jobs, and kill if need be.
We would prefer to softly terminate jobs so that they have a chance to
clean up before shutting down.
"""
for job in self._active_jobs:
job.proc.interrupt()
if verbose and self._currently_processed is not None:
print(textwrap.dedent(f"""
Failed when processing the following Job:
Label: {self._currently_processed.label}
AutoLabels: {self._currently_processed.autolabels}
Source cmd: {self._currently_processed.source_cmd}
""").strip() + "\n")
if self._active_jobs:
time.sleep(0.5)
remaining_jobs = [j for j in self._active_jobs if j.proc.poll() is None]
if remaining_jobs:
print(
f'SIGINT sent to {len(self._active_jobs)} jobs, '
f'{len(remaining_jobs)} have not yet exited.\n'
'Entering short cleanup loop, after which stragglers will '
'be forcibly terminated.'
)
for _ in range(5):
time.sleep(2.0)
remaining_jobs = [j for j in remaining_jobs if j.proc.poll() is None]
if remaining_jobs:
print(f'{len(remaining_jobs)} still remain.')
else:
print('All remaining jobs have gracefully terminated.')
return
print(f'{len(remaining_jobs)} jobs refused to exit. Forcibly terminating.')
for j in remaining_jobs:
j.proc.terminate()
def _canary_import(self) -> None:
"""Make sure we can import torch before launching a slew of workers."""
source_cmds: Set[str] = set()
for w in self._work_items:
if w.source_cmd is not None:
source_cmds.add(f"{w.source_cmd} && ")
for source_cmd in (source_cmds or {""}):
cmd = f'{source_cmd}{PYTHON_CMD} -c "import torch"'
proc = subprocess.run(
cmd,
shell=True,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
encoding="utf-8",
executable=SHELL,
)
if proc.returncode:
raise ImportError(
f'Failed to import torch in subprocess: {cmd}\n{proc.stdout}')
|