-
Notifications
You must be signed in to change notification settings - Fork 0
/
recursive_hd_reduce.py
154 lines (138 loc) · 6.22 KB
/
recursive_hd_reduce.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
import argparse
import torch
import logging
import time
import copy
from torch import distributed as dist
DEVICE = "cpu"
#TENSOR_SIZE = 4
send_time = 0.0
receive_time = 0.0
def init_process(master_ip, rank, world_size):
dist.init_process_group(backend="gloo",
init_method="tcp://" + master_ip + ":6585",
rank=rank,
world_size=world_size)
def main(rank, num_nodes, tensor_size):
# Create a random tensor
tensor = torch.rand(tensor_size)
#tensor = torch.arange(TENSOR_SIZE) # TBD: Easier to test with regular integers t = torch.rand(TENSOR_SIZE)
#tensor = torch.tensor([0,1,2,3])
#s = time.time()
#print("Split Tensor:",split_tensor)
#for j in range(len(split_tensor)):
#print("Rank",rank,"Before",tensor)
bde_reduce_scatter(rank, tensor, 0, len(tensor)-1, 0, num_nodes-1) # rank-1, rank+1)
#split_tensor = list(torch.split(tensor, int(tensor_size/num_nodes))) # Split tensor into chunks based on number of participating nodes
#print("Rank",rank,"After Reduce Scatter",tensor)
bde_all_gather(rank, tensor, 0, len(tensor)-1, 0 , num_nodes-1)
#print("Rank",rank,"After All Gather",tensor)
#split_tensor[rank] = bde_reduce_scatter(rank, tensor, 0, len(tensor)-1, 0, num_nodes-1) # rank-1, rank+1)
#print("x[",rank,"] = ",split_tensor[rank])
#e = time.time()
print("Rank",rank,"finished BDE all reduce in send_time", send_time, " seconds", "and receive_time", receive_time," seconds")
# BDE Reduce-Scatter
def bde_reduce_scatter(rank, x, tensor_left, tensor_right, rank_left, rank_right):
#print("Rank",rank,"x",x,"tensor_left",tensor_left,"tensor_right",tensor_right,"rank_left",rank_left,"rank_right",rank_right)
#print("Rank:",rank)
global send_time, receive_time
if rank_left == rank_right:
#val = x[(rank * TENSOR_SIZE)//dist.get_world_size():((rank + 1)*TENSOR_SIZE//dist.get_world_size())]
return
rank_size = rank_right - rank_left + 1
tensor_size = tensor_right - tensor_left + 1
tensor_mid = (tensor_left + tensor_right) // 2
rank_mid = (rank_left + rank_right) // 2
if rank <= rank_mid:
partner = rank + (rank_size/2)
else:
partner = rank - (rank_size/2)
partner = int(partner)
#print("Partner of",rank,"is",partner)
if rank <= rank_mid:
#print("Rank",rank,"is sending",x[tensor_mid+1:tensor_right+1],"of len",len(x[tensor_mid+1:tensor_right+1]))
ss = time.time()
dist.send(x[tensor_mid+1:tensor_right+1], dst=partner)
es = time.time()
send_time += es - ss
#print("Rank",rank,"Sending",x[tensor_mid+1:tensor_right+1])
recv_buffer = copy.deepcopy(x[tensor_left:tensor_mid+1]) #torch.zeros(len(x[tensor_left:tensor_mid+1]))
sr = time.time()
dist.recv(recv_buffer, src=partner)
er = time.time()
receive_time += er - sr
#print("Rank",rank,"Recv Buf",recv_buffer)
x[tensor_left:tensor_mid+1] = x[tensor_left:tensor_mid+1] + recv_buffer
else:
recv_buffer = copy.deepcopy(x[tensor_mid+1:tensor_right+1]) #torch.zeros(len(x[tensor_mid+1:tensor_right+1]))
#print("Rank",rank,"is expected to receive len",len(x[tensor_mid+1:tensor_right+1]))
sr = time.time()
dist.recv(recv_buffer, src=partner)
er = time.time()
receive_time += er - sr
#print("Rank",rank,"Recv Buf",recv_buffer)
x[tensor_mid+1:tensor_right+1] = x[tensor_mid+1:tensor_right+1] + recv_buffer
ss = time.time()
dist.send(x[tensor_left:tensor_mid+1], dst=partner)
es = time.time()
send_time += es - ss
#print("Rank",rank,"Sending",x[tensor_left:tensor_mid+1])
if rank <= rank_mid:
bde_reduce_scatter(rank, x, tensor_left, tensor_mid, rank_left, rank_mid)
else:
bde_reduce_scatter(rank, x, tensor_mid+1, tensor_right, rank_mid+1, rank_right)
# BDE AllGather
def bde_all_gather(rank, x, tensor_left, tensor_right, rank_left, rank_right):
global send_time, receive_time
#print("Rank",rank,"x",x,"tensor_left",tensor_left,"tensor_right",tensor_right,"rank_left",rank_left,"rank_right",rank_right)
if rank_left == rank_right:
return
rank_size = rank_right - rank_left + 1
tensor_size = tensor_right - tensor_left +1
tensor_mid = (tensor_left + tensor_right) // 2
rank_mid = (rank_left + rank_right) // 2
if rank <= rank_mid:
partner = rank + (rank_size/2)
else:
partner = rank - (rank_size/2)
partner = int(partner)
if rank <= rank_mid:
bde_all_gather(rank, x, tensor_left, tensor_mid, rank_left, rank_mid)
else:
bde_all_gather(rank, x, tensor_mid+1, tensor_right, rank_mid+1, rank_right)
if rank <= rank_mid:
#print("Rank",rank,"sending",x[tensor_left:tensor_mid+1],"from",x)
ss = time.time()
dist.send(x[tensor_left:tensor_mid+1], dst=partner)
es = time.time()
send_time += es - ss
recv_buffer = x[tensor_mid+1:tensor_right+1]
sr = time.time()
dist.recv(recv_buffer, src=partner)
er = time.time()
receive_time += er - sr
#print("Rank",rank,"receiving",recv_buffer)
else:
recv_buffer = x[tensor_left:tensor_mid+1]
sr = time.time()
dist.recv(recv_buffer, src=partner)
er = time.time()
receive_time += er - sr
#print("Rank",rank,"receiving",recv_buffer)
#print("Rank",rank,"sending",x[tensor_left:tensor_mid+1],"from",x)
ss = time.time()
dist.send(x[tensor_mid+1:tensor_right+1], dst=partner)
es = time.time()
send_time += es - ss
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--master-ip", "-m", required=True, type=str)
parser.add_argument("--num-nodes", "-n", required=True, type=int)
parser.add_argument("--rank", "-r", required=True, type=int)
parser.add_argument("--tensor-size", "-t", required=True, type=int)
args = parser.parse_args()
#print("Rank",args.rank,"entered main")
init_process(master_ip=args.master_ip,
rank=args.rank,
world_size=args.num_nodes)
main(rank=args.rank, num_nodes=args.num_nodes, tensor_size=args.tensor_size)