File size: 10,343 Bytes
256a159
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
259
260
261
262
263
264
265
import os
import os.path as osp
import re
import subprocess
import time
import traceback
from functools import partial
from multiprocessing import Pipe, Pool
from typing import Any, Dict, List, Optional, Tuple

import mmengine
from mmengine.config import ConfigDict
from tqdm import tqdm

from opencompass.registry import RUNNERS, TASKS
from opencompass.utils import batched, get_logger

from .base import BaseRunner


@RUNNERS.register_module()
class SlurmSequentialRunner(BaseRunner):
    """Distributed runner based on Slurm. It will launch tasks in parallel
    using `srun` command.

    This runner launches tasks one by one for execution. A new task will only
    be launched when and only when max_num_workers is not met, and the previous
    task has been successfully allocated to a machine. Therefore, unlike the
    `SlurmRunner`, at most only one task will be in the PENDING status at the
    same time during a run, making the random_sleep strategy no longer
    necessary. In addition, this runner also includes a feature to
    automatically kill all jobs by the job_id on exit.

    The runner will obtain the job_id by reading the srun output similar to
    `srun: Job 123456 scheduled successfully!`. If the output of srun does not
    match this pattern, the runner will not work properly.

    Args:
        task (ConfigDict): Task type config.
        max_num_workers (int): Max number of workers to run in parallel.
            Defaults to 32.
        retry (int): Number of retries if the job failed. Defaults to 2.
        partition (str): Slurm partition name. Defaults to None.
        quotatype (str): Slurm quota type. Defaults to None.
        qos (str): Slurm quality of service. Defaults to None.
        debug (bool): Whether to run in debug mode. Defaults to False.
        lark_bot_url (str): Lark bot url. Defaults to None.
        extra_command (List, optional): Extra slurm command.
            For example ['-c 12', '-w node1']. Defaults to None.
    """

    def __init__(self,
                 task: ConfigDict,
                 task_prefix: str = '',
                 max_num_workers: int = 32,
                 retry: int = 2,
                 partition: str = None,
                 quotatype: str = None,
                 qos: str = None,
                 debug: bool = False,
                 lark_bot_url: str = None,
                 extra_command: Optional[List[str]] = None):
        super().__init__(task=task, debug=debug, lark_bot_url=lark_bot_url)
        self.max_num_workers = max_num_workers
        self.retry = retry
        self.partition = partition
        self.quotatype = quotatype
        self.qos = qos
        self.task_prefix = task_prefix
        if not extra_command:
            extra_command = []
        assert isinstance(extra_command, list)
        self.extra_command = extra_command

        logger = get_logger()
        if self.quotatype in ['spot', 'auto']:
            logger.warning(
                'Quotatype spot or auto may cause stability issues, '
                'reserved is recommended.')

    def launch(self, tasks: List[Dict[str, Any]]) -> List[Tuple[str, int]]:
        if not self.debug:
            return self._launch_wo_debug(tasks)
        else:
            return [self._launch(task) for task in tasks]

    def _launch_wo_debug(self,
                         tasks: List[Dict[str, Any]]) -> List[Tuple[str, int]]:
        launched_bar = tqdm(total=len(tasks), desc='Launched')
        finished_bar = tqdm(total=len(tasks), desc='Finished')
        job_ids = []
        status = []

        def _update(result):
            finished_bar.update()
            status.append(result)
            return result

        def _err_update(err):
            finished_bar.update()
            traceback.print_exc()
            status.append(('', -1))

        try:
            parent_conns = []
            num_workers = max(min(self.max_num_workers, len(tasks)), 1)
            with Pool(processes=num_workers) as pool:
                for task in tasks:
                    parent_conn, child_conn = Pipe()
                    _ = pool.apply_async(self._launch,
                                         kwds={
                                             'cfg': task,
                                             'child_conn': child_conn
                                         },
                                         callback=_update,
                                         error_callback=_err_update)
                    time.sleep(0.5)

                    job_id = parent_conn.recv()
                    launched_bar.update()
                    parent_conns.append(parent_conn)
                    job_ids.append(job_id)

                pool.close()
                pool.join()
            return status
        except KeyboardInterrupt:
            raise
        finally:
            launched_bar.close()
            finished_bar.close()
            for parent_conn in parent_conns:
                while parent_conn.poll():
                    try:
                        job_id = parent_conn.recv()
                        job_ids.append(job_id)
                    except EOFError:
                        break
                parent_conn.close()

            tbar = tqdm(total=len(job_ids), desc='clear sruns')
            for batched_job_ids in batched(job_ids, 4):
                while True:
                    ps = []
                    try:
                        for job_id in batched_job_ids:
                            tbar.update()
                            if job_id is None:
                                continue
                            cmd = f'scancel {job_id}'
                            p = subprocess.Popen(cmd,
                                                 shell=True,
                                                 stdout=subprocess.PIPE,
                                                 stderr=subprocess.STDOUT)
                            ps.append(p)
                        break
                    except KeyboardInterrupt:
                        logger = get_logger()
                        logger.error('Ignoring KeyboardInterrupt...')
                for p in ps:
                    p.wait()
            tbar.close()

    def _launch(self, cfg: ConfigDict, child_conn: Pipe = None):
        logger = get_logger()

        task = TASKS.build(dict(cfg=cfg, type=self.task_cfg['type']))
        num_gpus = task.num_gpus
        task_name = task.name
        task_name = self.task_prefix + task_name

        # Dump task config to file
        mmengine.mkdir_or_exist('tmp/')
        param_file = f'tmp/{os.getpid()}_params.py'
        process = None
        try:
            cfg.dump(param_file)

            # Build up slurm command
            tmpl = 'srun'
            if self.partition:
                tmpl += f' -p {self.partition}'
            if self.quotatype:
                tmpl += f' --quotatype={self.quotatype}'
            if self.qos:
                tmpl += f' --qos={self.qos}'
            if num_gpus > 0:
                tmpl += f' --gres=gpu:{num_gpus}'
            for extra_cmd in self.extra_command:
                tmpl += f' {extra_cmd}'
            tmpl += f" -N1 -u -J '{task_name[:512]}'" + ' {task_cmd}'
            get_cmd = partial(task.get_command,
                              cfg_path=param_file,
                              template=tmpl)
            cmd = get_cmd()

            logger.debug(f'Running command: {cmd}')

            retry = self.retry
            output_paths = task.get_output_paths()

            if self.debug:
                while True:
                    process = subprocess.Popen(cmd, shell=True, text=True)
                    process.communicate()
                    process.wait()
                    if self._job_failed(process.returncode, output_paths):
                        if retry > 0:
                            logger.warning(
                                f'task {task_name} failed, retrying...')
                            retry -= 1
                            cmd = get_cmd()
                        else:
                            break
                    else:
                        break
            else:
                out_path = task.get_log_path(file_extension='out')
                mmengine.mkdir_or_exist(osp.split(out_path)[0])
                stdout = open(out_path, 'w', encoding='utf-8')
                stderr = subprocess.PIPE
                while True:
                    process = subprocess.Popen(cmd,
                                               shell=True,
                                               text=True,
                                               stdout=stdout,
                                               stderr=stderr)
                    job_id = None
                    while True:
                        line = process.stderr.readline()
                        if not line:
                            break
                        match = re.search(
                            r'srun: Job (\d+) scheduled successfully!', line)
                        if match and job_id is None:
                            job_id = match.group(1)
                            child_conn.send(job_id)
                        stdout.write(line)
                    process.wait()
                    if self._job_failed(process.returncode, output_paths):
                        if retry > 0:
                            retry -= 1
                            cmd = get_cmd()
                        else:
                            logger.warning(
                                f'task {task_name} fail, see\n{out_path}')
                            break
                    else:
                        break
        except KeyboardInterrupt:
            raise
        finally:
            # Clean up
            if child_conn is not None:
                child_conn.send(None)
                child_conn.close()
            if process is not None:
                process.kill()
            os.remove(param_file)
        return task_name, process.returncode

    def _job_failed(self, return_code: int, output_paths: List[str]) -> bool:
        return return_code != 0 or not all(
            osp.exists(output_path) for output_path in output_paths)