|
1
|
|
|
from __future__ import absolute_import |
|
2
|
|
|
|
|
3
|
|
|
import os |
|
4
|
|
|
import re |
|
5
|
|
|
from smartdispatch.pbs import PBS |
|
6
|
|
|
from smartdispatch import utils |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
def job_generator_factory(queue, commands, prolog=[], epilog=[], command_params={}, cluster_name=None, base_path="./"): |
|
10
|
|
|
if cluster_name == "guillimin": |
|
11
|
|
|
return GuilliminJobGenerator(queue, commands, prolog, epilog, command_params, base_path) |
|
12
|
|
|
elif cluster_name == "mammouth": |
|
13
|
|
|
return MammouthJobGenerator(queue, commands, prolog, epilog, command_params, base_path) |
|
14
|
|
|
elif cluster_name == "helios": |
|
15
|
|
|
return HeliosJobGenerator(queue, commands, prolog, epilog, command_params, base_path) |
|
16
|
|
|
elif cluster_name == "hades": |
|
17
|
|
|
return HadesJobGenerator(queue, commands, prolog, epilog, command_params, base_path) |
|
18
|
|
|
|
|
19
|
|
|
return JobGenerator(queue, commands, prolog, epilog, command_params, base_path) |
|
20
|
|
|
|
|
21
|
|
|
|
|
22
|
|
|
class JobGenerator(object): |
|
23
|
|
|
|
|
24
|
|
|
""" Offers functionalities to generate PBS files for a given queue. |
|
25
|
|
|
|
|
26
|
|
|
Parameters |
|
27
|
|
|
---------- |
|
28
|
|
|
queue : `Queue` instance |
|
29
|
|
|
queue on which commands will be executed |
|
30
|
|
|
commands : list of str |
|
31
|
|
|
commands to put in PBS files |
|
32
|
|
|
prolog : list of str |
|
33
|
|
|
code to execute before the commands |
|
34
|
|
|
epilog : list of str |
|
35
|
|
|
code to execute after the commands |
|
36
|
|
|
command_params : dict |
|
37
|
|
|
information about the commands |
|
38
|
|
|
""" |
|
39
|
|
|
|
|
40
|
|
|
def __init__(self, queue, commands, prolog=[], epilog=[], command_params={}, base_path="./"): |
|
41
|
|
|
self.prolog = prolog |
|
42
|
|
|
self.commands = commands |
|
43
|
|
|
self.epilog = epilog |
|
44
|
|
|
self.queue = queue |
|
45
|
|
|
self.job_log_filename = '"{base_path}/logs/job/"$PBS_JOBID".{{ext}}"'.format(base_path=base_path) |
|
46
|
|
|
|
|
47
|
|
|
self.nb_cores_per_command = command_params.get('nb_cores_per_command', 1) |
|
48
|
|
|
self.nb_gpus_per_command = command_params.get('nb_gpus_per_command', 1) |
|
49
|
|
|
#self.mem_per_command = command_params.get('mem_per_command', 0.0) |
|
50
|
|
|
|
|
51
|
|
|
self.pbs_list = self._generate_base_pbs() |
|
52
|
|
|
self._add_cluster_specific_rules() |
|
53
|
|
|
|
|
54
|
|
|
def _add_cluster_specific_rules(self): |
|
55
|
|
|
pass |
|
56
|
|
|
|
|
57
|
|
|
def add_pbs_flags(self, flags): |
|
58
|
|
|
resources = {} |
|
59
|
|
|
options = {} |
|
60
|
|
|
|
|
61
|
|
|
for flag in flags: |
|
62
|
|
|
flag = flag |
|
63
|
|
|
if flag.startswith('-l'): |
|
64
|
|
|
resource = flag[2:] |
|
65
|
|
|
split = resource.find('=') |
|
66
|
|
|
resources[resource[:split]] = resource[split+1:] |
|
67
|
|
|
elif flag.startswith('-'): |
|
68
|
|
|
options[flag[1:2]] = flag[2:] |
|
69
|
|
|
else: |
|
70
|
|
|
raise ValueError("Invalid PBS flag ({})".format(flag)) |
|
71
|
|
|
|
|
72
|
|
|
for pbs in self.pbs_list: |
|
73
|
|
|
pbs.add_resources(**resources) |
|
74
|
|
|
pbs.add_options(**options) |
|
75
|
|
|
|
|
76
|
|
|
def add_sbatch_flags(self, flags): |
|
77
|
|
|
options = {} |
|
78
|
|
|
|
|
79
|
|
|
for flag in flags: |
|
80
|
|
|
split = flag.find('=') |
|
81
|
|
|
options[flag[:split]] = flag[split+1:] |
|
82
|
|
|
|
|
83
|
|
|
for pbs in self.pbs_list: |
|
84
|
|
|
pbs.add_sbatch_options(**options) |
|
85
|
|
|
|
|
86
|
|
|
def _generate_base_pbs(self): |
|
87
|
|
|
""" Generates PBS files allowing the execution of every commands on the given queue. """ |
|
88
|
|
|
nb_commands_per_node = self.queue.nb_cores_per_node // self.nb_cores_per_command |
|
89
|
|
|
|
|
90
|
|
|
if self.queue.nb_gpus_per_node > 0 and self.nb_gpus_per_command > 0: |
|
91
|
|
|
nb_commands_per_node = min(nb_commands_per_node, self.queue.nb_gpus_per_node // self.nb_gpus_per_command) |
|
92
|
|
|
|
|
93
|
|
|
pbs_files = [] |
|
94
|
|
|
# Distribute equally the jobs among the PBS files and generate those files |
|
95
|
|
|
for i, commands in enumerate(utils.chunks(self.commands, n=nb_commands_per_node)): |
|
96
|
|
|
pbs = PBS(self.queue.name, self.queue.walltime) |
|
97
|
|
|
|
|
98
|
|
|
# TODO Move the add_options into the JobManager once created. |
|
99
|
|
|
pbs.add_options(o=self.job_log_filename.format(ext='out'), e=self.job_log_filename.format(ext='err')) |
|
100
|
|
|
|
|
101
|
|
|
# Set resource: nodes |
|
102
|
|
|
resource = "1:ppn={ppn}".format(ppn=len(commands) * self.nb_cores_per_command) |
|
103
|
|
|
if self.queue.nb_gpus_per_node > 0: |
|
104
|
|
|
resource += ":gpus={gpus}".format(gpus=len(commands) * self.nb_gpus_per_command) |
|
105
|
|
|
|
|
106
|
|
|
pbs.add_resources(nodes=resource) |
|
107
|
|
|
|
|
108
|
|
|
pbs.add_modules_to_load(*self.queue.modules) |
|
109
|
|
|
pbs.add_to_prolog(*self.prolog) |
|
110
|
|
|
pbs.add_commands(*commands) |
|
111
|
|
|
pbs.add_to_epilog(*self.epilog) |
|
112
|
|
|
|
|
113
|
|
|
pbs_files.append(pbs) |
|
114
|
|
|
|
|
115
|
|
|
return pbs_files |
|
116
|
|
|
|
|
117
|
|
|
def write_pbs_files(self, pbs_dir="./"): |
|
118
|
|
|
""" Writes PBS files allowing the execution of every commands on the given queue. |
|
119
|
|
|
|
|
120
|
|
|
Parameters |
|
121
|
|
|
---------- |
|
122
|
|
|
pbs_dir : str |
|
123
|
|
|
folder where to save pbs files |
|
124
|
|
|
""" |
|
125
|
|
|
pbs_filenames = [] |
|
126
|
|
|
for i, pbs in enumerate(self.pbs_list): |
|
127
|
|
|
pbs_filename = os.path.join(pbs_dir, 'job_commands_' + str(i) + '.sh') |
|
128
|
|
|
pbs.save(pbs_filename) |
|
129
|
|
|
pbs_filenames.append(pbs_filename) |
|
130
|
|
|
|
|
131
|
|
|
return pbs_filenames |
|
132
|
|
|
|
|
133
|
|
|
def specify_account_name_from_env(self, environment_variable_name): |
|
134
|
|
|
if environment_variable_name not in os.environ: |
|
135
|
|
|
raise ValueError("Undefined environment variable: ${}. Please, provide your account name!".format(environment_variable_name)) |
|
136
|
|
|
|
|
137
|
|
|
account_name = os.path.basename(os.path.realpath(os.getenv(environment_variable_name))) |
|
138
|
|
|
for pbs in self.pbs_list: |
|
139
|
|
|
pbs.add_options(A=account_name) |
|
140
|
|
|
|
|
141
|
|
|
def specify_account_name_from_file(self, rapid_filename): |
|
142
|
|
|
if not os.path.isfile(rapid_filename): |
|
143
|
|
|
raise ValueError("Account name file {} does not exist. Please, provide your account name!".format(rapid_filename)) |
|
144
|
|
|
|
|
145
|
|
|
with open(rapid_filename, 'r') as rapid_file: |
|
146
|
|
|
account_name = rapid_file.read().strip() |
|
147
|
|
|
|
|
148
|
|
|
for pbs in self.pbs_list: |
|
149
|
|
|
pbs.add_options(A=account_name) |
|
150
|
|
|
|
|
151
|
|
|
|
|
152
|
|
|
class MammouthJobGenerator(JobGenerator): |
|
153
|
|
|
|
|
154
|
|
|
def _add_cluster_specific_rules(self): |
|
155
|
|
|
if self.queue.name.endswith("@mp2"): |
|
156
|
|
|
for pbs in self.pbs_list: |
|
157
|
|
|
pbs.resources['nodes'] = re.sub("ppn=[0-9]+", "ppn=1", pbs.resources['nodes']) |
|
158
|
|
|
|
|
159
|
|
|
|
|
160
|
|
|
class HadesJobGenerator(JobGenerator): |
|
161
|
|
|
|
|
162
|
|
|
def _add_cluster_specific_rules(self): |
|
163
|
|
|
for pbs in self.pbs_list: |
|
164
|
|
|
gpus = re.match(".*gpus=([0-9]+)", pbs.resources['nodes']).group(1) |
|
165
|
|
|
pbs.resources['nodes'] = re.sub("ppn=[0-9]+", "ppn={}".format(gpus), pbs.resources['nodes']) |
|
166
|
|
|
pbs.resources['nodes'] = re.sub(":gpus=[0-9]+", "", pbs.resources['nodes']) |
|
167
|
|
|
|
|
168
|
|
|
|
|
169
|
|
|
class GuilliminJobGenerator(JobGenerator): |
|
170
|
|
|
|
|
171
|
|
|
def _add_cluster_specific_rules(self): |
|
172
|
|
|
return self.specify_account_name_from_env('HOME_GROUP') |
|
173
|
|
|
|
|
174
|
|
|
|
|
175
|
|
|
# https://wiki.calculquebec.ca/w/Ex%C3%A9cuter_une_t%C3%A2che#tab=tab6 |
|
176
|
|
|
class HeliosJobGenerator(JobGenerator): |
|
177
|
|
|
|
|
178
|
|
|
def _add_cluster_specific_rules(self): |
|
179
|
|
|
self.specify_account_name_from_file(os.path.join(os.environ['HOME'], ".default_rap")) |
|
180
|
|
|
|
|
181
|
|
|
for pbs in self.pbs_list: |
|
182
|
|
|
# Remove forbidden ppn option. Default is 2 cores per gpu. |
|
183
|
|
|
pbs.resources['nodes'] = re.sub(":ppn=[0-9]+", "", pbs.resources['nodes']) |
|
184
|
|
|
|
|
185
|
|
|
class SlurmClusterGenerator(JobGenerator): |
|
186
|
|
|
|
|
187
|
|
|
def _add_cluster_specific_rules(self): |
|
188
|
|
|
for pbs in self.pbs_list: |
|
189
|
|
|
node_resource = pbs.resources.pop('nodes') |
|
190
|
|
|
gpus = re.match(".*gpus=([0-9]+)", node_resource).group(1) |
|
191
|
|
|
ppn = re.match(".*ppn=([0-9]+)", node_resource).group(1) |
|
192
|
|
|
pbs.add_resources(naccelerators=gpus) |
|
193
|
|
|
pbs.add_resources(ncpus=ppn) |
|
194
|
|
|
|
SMART-Lab /
smartdispatch
