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 | class PipelineMPI:
"""PipelineMPI class, it orchestrates the instantiation of MPI jobs
and distributes the workload among the workers.
"""
def __init__(
self,
exec: callable = None,
extra_params: dict = None,
collect: bool = None,
show_log: bool = True,
) -> None:
self.exec = exec
self.show_log = show_log
if extra_params is not None:
self.extra_params = extra_params
else:
self.extra_params = {}
self.collect = collect
self.comm = MPI.COMM_WORLD
self.n_procs = self.comm.Get_size()
self.status = (self.n_procs - 1) * [False]
self.status_dict = dict()
def _check_kwargs_consistency(self, kwargs: dict = None) -> int:
"""It checks if the kwargs provided for each worker
have the same length.
Args:
kwargs (dict, optional): a dictionary containing the kwargs of all (Default value = None)
Returns:
int: Length of the batch sent for each worker.
"""
types = [type(value) for value in kwargs.values()]
lengths = [len(value) for value in kwargs.values()]
assert all([t == list for t in types]), (
f"All the elements in kwargs must be list," f" but received {types}."
)
assert len(set(lengths)) == 1, (
f"All the elements in kwargs must be the same length,"
f" but received {lengths}"
)
print("kwargs is alright.")
return lengths[0]
def _split_kwargs(
self, kwargs: dict, rank: int, size: int, total_size: int
) -> Tuple[dict, int]:
"""It allows the workload be executed serially in each worker node
Args:
kwargs (dict): A dictionary containing kwargs, which will be distributed for all the workers.
rank (int): The index of the rank.
size (int): The number of available workers.
total_size (int): The total number of elements to be distributed among the workers.
Returns:
kwargs_batch: A dictionary containing the kwargs to be sent for each worker.
batch_size: The batch size, which corresponds to the number of elements
to be sent for each worker.
"""
# Decrement rank and size by 1, because they are usually 0-indexed in Python
size -= 1
rank -= 1
# Calculate batch size and remainder using divmod() function
batch_size, remainder = divmod(total_size, size)
# If rank is less than remainder, calculate kwargs_batch using batch size + 1
if rank < remainder:
kwargs_batch = {
key: value[rank * (batch_size + 1) : (rank + 1) * (batch_size + 1)]
for key, value in kwargs.items()
}
return kwargs_batch, batch_size + 1
# If rank is not less than remainder, calculate kwargs_batch using batch size
else:
kwargs_batch = {
key: value[
remainder * (batch_size + 1)
+ (rank - remainder)
* batch_size : (rank - remainder + 1)
* batch_size
]
for key, value in kwargs.items()
}
return kwargs_batch, batch_size
def _attribute_dict_output(self, dicts: list = None) -> None:
root = dict()
for e in dicts:
root.update(e)
for key, value in root.items():
self.status_dict[key] = value
@staticmethod
def inner_type(obj: list = None):
types_list = [type(o) for o in obj]
assert len(set(types_list)) == 1, "Composed types are not supported."
return types_list[0]
def _exec_wrapper(self, kwargs: dict, total_size: int) -> None:
"""A wrapper method around exec to facilitate the
instantiation of each worker.
Args:
kwargs (dict): A dictionary containing kwargs for the worker.
total_size (int): The total number of elements.
"""
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
size = comm.Get_size()
size_ = size
# Rank 0 is the 'master' node
# The worker nodes execute their workload and send a message to
# master
if rank != 0:
print(f"Executing rank {rank}.")
kwargs_batch, batch_size = self._split_kwargs(
kwargs, rank, size_, total_size
)
kwargs_batch_list = [
{key: value[j] for key, value in kwargs_batch.items()}
for j in range(batch_size)
]
out = list()
for i in kwargs_batch_list:
print(f"Executing batch {i['key']} in rank {rank}")
# Concatenate the rank to the extra parameters
i.update(self.extra_params)
# Appending the result of the operation self.exec to the partial list
out.append(self.exec(**i))
if self.collect is True:
msg = out
else:
msg = 1
if self.show_log:
print(f"Sending the output {msg} to rank 0")
comm.send(msg, dest=0)
print(f"Execution concluded for rank {rank}.")
# The master awaits the responses of each worker node
elif rank == 0:
for r in range(1, size):
msg = comm.recv(source=r)
self.status[r - 1] = msg
if self.inner_type(msg) == dict:
self._attribute_dict_output(dicts=msg)
if self.show_log:
print(f"Rank 0 received {msg} from rank {r}")
comm.barrier()
@property
def success(self) -> bool:
"""It returns True if the entire process worked without issues.
"""
return all(self.status)
def run(self, kwargs: dict = None) -> None:
"""It runs the MPI job
Args:
kwargs (dict, optional): A kwargs dictionary containing chunks of input arguments to be sent for each worker. (Default value = None)
"""
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
total_size = 0
# Checking if the datasets dimensions are in accordance with the expected ones
if rank == 0:
total_size = self._check_kwargs_consistency(kwargs=kwargs)
total_size = comm.bcast(total_size, root=0)
comm.barrier()
# Executing a wrapper containing the parallelized operation
self._exec_wrapper(kwargs, total_size)
comm.barrier()
|