tuning_comparison.py
4.88 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
160
161
162
import os,sys,argparse
import json
import itertools
from collections import namedtuple
from matplotlib import pyplot as plt
from common import *
def within_threshold(value, base, threshold):
bmax = base * ((100 + threshold)/100.0)
if value <= bmax:
return True
else:
return False
def find_optimal(data_points_list, inefficiency, ineff_threshold, cluster_threshold):
frontier = []
inefficiencies = []
energies = []
# Find emin
emin = min(data_points_list, key=lambda x: x['energy'])['energy']
# Check
for point in data_points_list:
assert float(point['energy']) >= emin
# Find all inefficiencies
for point in data_points_list:
inefficiencies.append(float(point['energy']) / float(emin))
# Find the frontier
for idx, ineff in enumerate(inefficiencies):
if within_threshold(ineff, inefficiency, ineff_threshold):
frontier.append(data_points_list[idx])
# Find the optimal point - point with best performance
optimal_point = min(frontier, key=lambda x: x['performance'])
# Check
for point in frontier:
assert point['performance'] >= optimal_point['performance']
return optimal_point, optimal_point['performance'], optimal_point['energy']
def no_tuning(args, bmark, inefficiency, ineff_threshold=0.0, cluster_threshold=0.0):
file_path = os.path.join(os.path.join(os.path.join(args.input_dir, "aggr_data"), bmark), 'frontiers.json')
data = json.loads(open(file_path).read())
nt_point, performance, energy = find_optimal(data['data'], inefficiency, ineff_threshold, cluster_threshold)
# Get the simulation corresponding to the optimal
file_path = os.path.join(os.path.join(os.path.join(args.input_dir, "per_sample_data"), bmark), 'per_sample_frontiers.json')
data = json.loads(open(file_path).read())
no_tuning_data = []
for point in data['data']:
for sample in point:
if sample['cpu_freq'] == nt_point['cpu_freq'] and sample['mem_freq'] == nt_point['mem_freq']:
no_tuning_data.append(sample)
assert len(no_tuning_data) == len(data['data'])
return nt_point, no_tuning_data, performance, energy
def with_tuning(args, bmark, inefficiency, ineff_threshold=0.0, cluster_threshold=0.0):
file_path = os.path.join(os.path.join(os.path.join(args.input_dir, "per_sample_data"), bmark), 'per_sample_frontiers.json')
data = json.loads(open(file_path).read())
optimal_points = []
performance = 0
energy = 0
# For each sample
for point in data['data']:
assert len(point) == len(data['data'][0]) # 496
optimal_point, performance, energy = find_optimal(point, inefficiency, ineff_threshold, cluster_threshold)
optimal_points.append(optimal_point)
return optimal_points, sum(x['performance'] for x in optimal_points), sum(x['energy'] for x in optimal_points)
def plot(args, bmark, inefficiency, nt_points_list, t_points_list):
fig, axes_array = plt.subplots(3, sharex=True)
x_axis = range(len(t_points_list[0]))
x_ticklabels = [str(x) for x in x_axis]
ax = axes_array[0]
ax.set_ylabel('Inefficiency')
for nt_points, t_points in itertools.izip(nt_points_list, t_points_list):
nt_ineff = [x['inefficiency'] for x in nt_points]
t_ineff = [x['inefficiency'] for x in t_points]
ax.plot(x_axis, nt_ineff, 'k', label='No Tuning')
ax.plot(x_axis, t_ineff, label='W. Tuning')
ax.legend()
ax = axes_array[1]
ax.set_ylabel('Performance (ms)')
for nt_points, t_points in itertools.izip(nt_points_list, t_points_list):
nt_perf = [x['performance']/1e6 for x in nt_points]
t_perf = [x['performance']/1e6 for x in t_points]
ax.plot(x_axis, nt_perf, 'k')
ax.plot(x_axis, t_perf)
ax = axes_array[2]
cpi = get_cpi(bmark, args.input_dir)
ax.set_xlabel('Instructions (x10 Million)')
ax.set_ylabel('CPI')
ax.plot(x_axis, cpi, 'g')
plt.tight_layout()
fig.subplots_adjust(hspace=0.03)
for ax in axes_array:
ax.grid(True)
ax.set_axisbelow(True)
out_fname = os.path.join(os.path.join(os.path.join(args.output_dir, 'tuning_comparison'), bmark), '%.1f.jpg' % (inefficiency))
if not os.path.exists(os.path.dirname(out_fname)):
os.makedirs(os.path.dirname(out_fname))
plt.savefig(out_fname)
def main(argv):
args = parse(argv)
bmarks, labels = get_benchmarks(args)
for bmark in bmarks:
print 'BENCHMARK: %s' % (bmark)
nt_points_list = []
t_points_list = []
for ineff in args.inefficiency:
string = '\tinefficiency: %0.2f\n' % (ineff)
nt_point, data, nt_performance, nt_energy = no_tuning(args, bmark, ineff, 3)
string += '\t\tNO TUNING: %0.2fms %0.2fmJ\n' % (nt_performance / 1e6, nt_energy / 1e6)
t_points, t_performance, t_energy = with_tuning(args, bmark, ineff, 3)
string += '\t\tW. TUNING: %0.2fms %0.2fmJ\n' % (t_performance / 1e6, t_energy / 1e6)
nt_points_list.append(data)
t_points_list.append(t_points)
print string
plot(args, bmark, ineff, [data], [t_points])
# Uncomment this line if you want an aggregate plot per benchmark (Not very clean)
#plot(args, bmark, nt_points_list, t_points_list)
if __name__ == '__main__':
main(sys.argv)