-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathpytorch_send_recv_test.py
156 lines (136 loc) · 6.23 KB
/
pytorch_send_recv_test.py
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
import random
import torch
import argparse
import time
import torch.distributed as dist
def test_sync_send_recv_delay(args, device, communicator):
print("<==== Test delay ====>")
if args.rank == 1:
send_tensor = torch.ones(1, dtype=torch.float32, device=device)
send_tensor += random.random()
if args.dist_backend == 'nccl':
dist.barrier(device_ids=[args.cuda_id])
torch.cuda.synchronize()
else:
dist.barrier()
if args.use_cuda:
torch.cuda.synchronize()
start_time = time.perf_counter_ns()
communicator.send(send_tensor, dst=0)
if args.use_cuda:
torch.cuda.synchronize()
end_time = time.perf_counter_ns()
estimated_delay = 1e-9 * (end_time - start_time)/2
print('Send tensor is done: estimated delay:', estimated_delay * 1000, "ms.")
elif args.rank == 0:
recv_tensor = torch.zeros(1, dtype=torch.float32, device=device)
if args.dist_backend == 'nccl':
dist.barrier(device_ids=[args.cuda_id])
else:
dist.barrier()
if args.use_cuda:
torch.cuda.synchronize()
start_time = time.perf_counter_ns()
communicator.recv(recv_tensor, src=1)
if args.use_cuda:
torch.cuda.synchronize()
end_time = time.perf_counter_ns()
estimated_delay = 1e-9 * (end_time - start_time)/2
print('Recv tensor is done: estimated delay:', estimated_delay * 1000, "ms.")
recv_tensor += random.random()
if args.use_cuda:
torch.cuda.synchronize()
return estimated_delay
def test_sync_send_recv_bandwidth(args, device, communicator):
print("<==== Test bandwidth ====>")
if args.rank == 1:
send_tensor = torch.arange(args.dim, dtype=torch.float32, device=device)
if args.dist_backend == 'nccl':
dist.barrier(device_ids=[args.cuda_id])
else:
dist.barrier()
if args.use_cuda:
torch.cuda.synchronize()
start_time = time.perf_counter_ns()
communicator.send(send_tensor, dst=0)
if args.use_cuda:
torch.cuda.synchronize()
end_time = time.perf_counter_ns()
total_time = 1e-9 * (end_time - start_time)
estimated_bandwidth = 8 * 4 * args.dim / total_time / 1024 / 1024 / 1024
print('Send tensor is done: tensor size:<', args.dim, "> takes:", total_time, "second, estimated bandwidth:",
estimated_bandwidth, "Gbps.")
elif args.rank == 0:
recv_tensor = torch.zeros(args.dim, dtype=torch.float32, device=device)
if args.dist_backend == 'nccl':
dist.barrier(device_ids=[args.cuda_id])
else:
dist.barrier()
if args.use_cuda:
torch.cuda.synchronize()
start_time = time.perf_counter_ns()
communicator.recv(recv_tensor, src=1)
if args.use_cuda:
torch.cuda.synchronize()
end_time = time.perf_counter_ns()
total_time = 1e-9 * (end_time - start_time)
estimated_bandwidth = 8 * 4 * args.dim / total_time / 1024 / 1024 / 1024
print('Recv tensor is done: tensor size:<', args.dim, "> takes:", total_time, "second, estimated bandwidth:",
estimated_bandwidth, "Gbps.")
print(recv_tensor[args.dim//2]==args.dim//2)
print(recv_tensor[args.dim-1]==args.dim-1)
recv_tensor += random.random()
if args.use_cuda:
torch.cuda.synchronize()
return estimated_bandwidth, total_time
def main():
parser = argparse.ArgumentParser(description='Test PyTorch Distributed')
parser.add_argument('--dist-backend', type=str, default='cupy_nccl', metavar='S',
help='PyTorch backend type')
parser.add_argument('--dist-url', type=str, default='tcp://127.0.0.1:9000', metavar='S',
help='master ip for distributed PyTorch')
parser.add_argument('--world-size', type=int, default=2, metavar='D',
help='world size (default: 2)')
parser.add_argument('--rank', type=int, default=0, metavar='R',
help='rank for distributed PyTorch')
parser.add_argument('--dim', type=int, default=4*2048*2048, metavar='R',
help='size of the tensor to be sent.') # this is an approximated size of a macro-bench
parser.add_argument('--use-cuda', default=True, type=lambda x: (str(x).lower() == 'true'),
help='if this is set to True, will use cuda to train')
parser.add_argument('--cuda-id', type=int, default=0, metavar='N',
help='cuda index, if the instance has multiple GPUs.')
parser.add_argument('--iter', type=int, default=20, metavar='R',
help='number of iterations for benchmark.')
args = parser.parse_args()
if args.use_cuda:
assert (torch.cuda.is_available())
device = torch.device('cuda', args.cuda_id)
else:
device = torch.device('cpu')
if args.dist_backend == 'cupy_nccl':
from nccl_backend import NCCLCommunicator
communicator = NCCLCommunicator(rank=args.rank, intra_gpu_rank=args.cuda_id,
world_size=args.world_size, master_ip=args.dist_url)
else:
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
rank=args.rank, world_size=args.world_size)
communicator = dist
estimated_bandwidth = 0
e2e_time = 0
for i in range(args.iter):
current_bandwidth, current_time = test_sync_send_recv_bandwidth(args, device, communicator)
estimated_bandwidth += current_bandwidth
e2e_time += current_time
time.sleep(1)
estimated_bandwidth /= args.iter
e2e_time /= args.iter
print("This is Rank-", args.rank, "see the result in Rank 0 node.")
if args.rank == 0:
print("This is the right result (recv side):")
else:
print("Record the result in the other side !!!!!!!!!")
print("<=====Averaged estimated bandwidth: ", estimated_bandwidth, "Gbps=====>")
print("<=====Averaged end to end time: ", e2e_time, "s for sending <", 4 * args.dim / 1024/1024,
"> MB data=====>")
if __name__ == '__main__':
main()