Repository navigation
Expand file tree
/
Copy pathparallelism.py
More file actions
160 lines (137 loc) · 7.25 KB
/
Copy pathparallelism.py
File metadata and controls
160 lines (137 loc) · 7.25 KB
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
import config
import util
import math
class Parallelism():
def __init__(self, exp_config):
self.autoPar = exp_config.sch_config.auto
self.lp = exp_config.sch_config.lp
self.mb = exp_config.sch_config.mb
self.kp_hidden_dim1 = exp_config.sch_config.kp_hidden_dim1
self.kp_softmax_dim1 = exp_config.sch_config.kp_softmax_dim1
self.kp_embedding_dim1 = exp_config.sch_config.kp_embedding_dim1
self.kp_projection_dim1 = exp_config.sch_config.kp_projection_dim1
self.kp_hidden_dim2 = exp_config.sch_config.kp_hidden_dim2
self.kp_softmax_dim2 = exp_config.sch_config.kp_softmax_dim2
self.kp_embedding_dim2 = exp_config.sch_config.kp_embedding_dim2
self.kp_projection_dim2 = exp_config.sch_config.kp_projection_dim2
self.dp = exp_config.sch_config.dp
self.kp_hidden_type = exp_config.sch_config.kp_hidden_type #1: CR, 2: RC
self.kp_softmax_type = exp_config.sch_config.kp_softmax_type #1: CR, 2: RC
self.kp_embedding_type = exp_config.sch_config.kp_embedding_type #1: CR, 2: RC
self.kp_projection_type = exp_config.sch_config.kp_projection_type #1: CR, 2: RC
self.exp_config = exp_config
self.t = exp_config.sch_config.t # type of parallelism, e.g., "CR", "RC", "none"
self.tp = exp_config.sch_config.tp
self.cp = exp_config.sch_config.cp
self.tp_sp = exp_config.sch_config.tp_sp
self.kp1 = exp_config.sch_config.kp1 # first parallelism parameter
self.kp2 = exp_config.sch_config.kp2 # second parallelism parameter
def findParallelStrategy(self):
if (self.autoPar == None or self.autoPar == False):
pass
elif (self.autoPar == "greedy"):
self.greedyScheduler()
elif (self.autoPar=="dynamic"):#dynamic programming solution
self.dynamic()
else:
print("Scheduling types supported are None/False/greedy/dynamic")
return NotImplemented
def dynamic(self):
print("Dynamic Programming Scheduling Not Implemented yet!")
return NotImplemented
def greedyScheduler(self):
#Find minimum number of parallel workers based on
#application's memory footprint and
#memory capacity per accelerator
#Step 1. Try to fit everything on one GPU
#Step 2. If it does not fit, check if individual items fit on one GPU,
#if it does allocate a seperate GPU per item
#Step 3. If it does not, try layer parallelism across hidden layers
#Step 4. If it does not, try kernel parallelism across hidden layers
#Step 5. For softmax and embedding layers, try kernel parallelism
tot_mem, embedding_mem, hidden_mem, softmax_mem, projection_mem, wt_mem, act_mem, point_mem = util.getTotMemReq(self.exp_config)
#tot_mem, embedding_mem, hidden_mem, softmax_mem, projection_mem = util.getTotMemReq(self.exp_config)
print("Value of M is {.1d}\n"
.format(self.M))
if (tot_mem < self.M):
self.lp = 1
self.kp_hidden_type = -1
self.kp_softmax_type = -1
self.kp_embedding_type = -1
self.kp_projection_type = -1
else:
self.kp_hidden_dim1 = 1
self.kp_hidden_dim2 = 1
self.kp_softmax_dim1 = 1
self.kp_softmax_dim2 = 1
self.kp_projection_dim1 = 1
self.kp_projection_dim2 = 1
L = self.L
lp = math.ceil(hidden_mem / M)
self.lp = (L if lp > L else lp)
self.kp_hidden = (1 if lp <= L else math.ceil(hidden_mem / L / M))
self.kp_softmax = math.ceil(softmax_mem / M)
self.kp_embedding = math.ceil(embedding_mem / M)
self.kp_projection = math.ceil(projection_mem / M)
self.findlp()
if (self.kp_hidden == 1):
self.kp_hidden_type = -1
if (self.kp_softmax == 1):
self.kp_softmax_type = -1
if (self.kp_projection == 1):
self.kp_projection_type = -1
if (self.kp_embedding == 1):
self.kp_embedding_type = -1
if self.kp_hidden_type == 1:
self.kp_hidden_dim1 = self.kp_hidden
self.kp_hidden_dim2 = 1
elif self.kp_hidden_type == 2:
#This is an arbiotrary choice until I have more insight
self.kp_hidden_dim1 = self.findDiv(math.ceil(math.sqrt(self.kp_hidden)), self.miniB)
self.kp_hidden_dim2 = self.kp_hidden_dim1
if self.kp_softmax_type == 1:
self.kp_softmax_dim1 = self.kp_softmax
self.kp_softmax_dim2 = 1
elif self.kp_softmax_type == 2:
#This is an arbitrary choice until I have more insight
self.kp_softmax_dim1 = self.findDiv(math.ceil(math.sqrt(self.kp_softmax)), self.miniB * self.S)
self.kp_softmax_dim2 = self.kp_softmax_dim1
if self.kp_projection_type == 1:
self.kp_projection_dim1 = self.kp_projection
self.kp_projection_dim2 = 1
elif self.kp_projection_type == 2:
#This is an arbitrary choice until I have more insight
self.kp_projection_dim1 = self.findDiv(math.ceil(math.sqrt(self.kp_projection)), self.miniB * self.S)
self.kp_projection_dim2 = self.kp_projection_dim1
#################Helper functions################
def bag(self, bag):
#Not implemented, backpacking stratgey
return 1
#Find how to use minimum number of layers for layer parallelism
def findlp(self):
tot_mem, embedding_mem, hidden_mem, softmax_mem, projection_mem = self.getTotMemReq()
self.M = self.mem_size
#If Kernel parallelism is one for all components then finding how to group
#layers together is a knap-sack problem with some dependency constraints st.
#neighbouring layers are bagged together
if (self.lp == 1 and self.kp_hidden == 1 and self.kp_embedding == 1 and
self.kp_projection == 1 and self.kp_softmax == 1 and tot_mem > M):
value = [embedding_mem, hidden_mem, projection_mem, softmax_mem]
num_bags = self.bag(value)
elif (self.lp > 1 and self.kp_hidden == 1 and self.kp_embedding == 1 and
self.kp_projection == 1 and self.kp_softmax == 1 and tot_mem > M):
value1 = [embedding_mem, hidden_mem / self.lp]
value2 = [hidden_mem / self.lp, projection_mem, softmax_mem]
num_bags1 = self.bag(value1)
num_bags2 = self.bag(value2)
num_bags = num_bags1 + num_bags2
self.lp = num_bags
#Find minimum value between A and B that B is divisible by
def findDiv(self, A, B):
smallestFactor = -1
for i in range(A,B+1):
if (B % i == 0):
smallestFactor = i
break
assert(smallestFactor != -1)
return smallestFactor