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
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
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
|
"""
Utilities for parameter extraction from data layout.
Parameters include the amount of keys, length of strings (both keys and values),
length of lists, ane more.
"""
from protocol_benchmarks import codegen_for_lib
import numpy as np
import ubjson
def _string_value_length(json):
if type(json) == str:
return len(json)
if type(json) == dict:
return sum(map(_string_value_length, json.values()))
if type(json) == list:
return sum(map(_string_value_length, json))
return 0
# TODO distinguish between int and uint, which is not visible from the
# data value alone
def _int_value_length(json):
if type(json) == int:
if json < 256:
return 1
if json < 65536:
return 2
return 4
if type(json) == dict:
return sum(map(_int_value_length, json.values()))
if type(json) == list:
return sum(map(_int_value_length, json))
return 0
def _string_key_length(json):
if type(json) == dict:
return sum(map(len, json.keys())) + sum(map(_string_key_length, json.values()))
return 0
def _num_keys(json):
if type(json) == dict:
return len(json.keys()) + sum(map(_num_keys, json.values()))
return 0
def _num_of_type(json, wanted_type):
ret = 0
if type(json) == wanted_type:
ret = 1
if type(json) == dict:
ret += sum(map(lambda x: _num_of_type(x, wanted_type), json.values()))
if type(json) == list:
ret += sum(map(lambda x: _num_of_type(x, wanted_type), json))
return ret
def json_to_param(json):
"""Return numeric parameters describing the structure of JSON data."""
ret = dict()
ret['strlen_keys'] = _string_key_length(json)
ret['strlen_values'] = _string_value_length(json)
ret['bytelen_int'] = _int_value_length(json)
ret['num_int'] = _num_of_type(json, int)
ret['num_float'] = _num_of_type(json, float)
ret['num_str'] = _num_of_type(json, str)
return ret
class Protolog:
"""
Loader and postprocessor for raw protobench (protocol-modeling/benchmark.py) data.
Converts data sorted by (arch,lib)/benchmark/index/attribute
to data sorted by (benchmark,index)/arch/lib/attribute.
Once constructed, a class object provides three members:
libraries -- array of library:config elements found in the benchmark results
architectures -- array of multipass architecture names
aggregate -- enriched log data, ordered by benchmark: {
('benchmark name', 'sub-benchmark index') : {
'architecture' : {
'library:options' : {
'attribute' : value (usually int or array)
}
}
}
}
aggregate attributes:
bss_{nop,ser,serdes} : whole-program Block Storage Segment (BSS) size
callcycles_raw : { 'C++ statement' : [CPU cycles for execution] ... }.
Not adjusted for 'nop' cycles -> values are a few cycles higher than true duration
cycles_{ser,des,enc,dec,encser,desdec} : cycles for complete (de)serialization step,
measured using just one counter start/stop (not a sum of callcycles_raw entries).
Adjusted for 'nop' cycles -> should give accurate function call duration
data_{secnop,ser,serdes} : whole-program Data Segment size
heap_{ser,des} : Maximum heap usage during step
serialized_size : Size (Bytes) of serialized data
stack_alloc_{ser,des} : Maximum stack usage (Bytes) during step.
Based on online analysis (comparison of memory dumps)
stack_set_{ser,des} : Number of stack bytes modified during step.
Based on online analysis (comparison of memory dumps), should be
smaller than the corresponding stack_alloc_ value
text_{nop,ser,serdes} : whole-program Text Segment (code/Flash) size
"""
def _median_cycles(data, key):
# There should always be more than just one measurement -- otherwise
# something went wrong
if len(data[key]) <= 1:
return np.nan
for val in data[key]:
# bogus data
if val > 10_000_000:
return np.nan
for val in data['nop']:
# bogus data
if val > 10_000_000:
return np.nan
# All measurements in data[key] cover the same instructions, so they
# should be identical.
return max(0, int(np.median(data[key][1:]) - np.median(data['nop'][1:])))
idem = lambda x: x
datamap = [
['bss_nop', 'bss_size_nop', idem],
['bss_ser', 'bss_size_ser', idem],
['bss_serdes', 'bss_size_serdes', idem],
['callcycles_raw', 'callcycles', idem],
['cycles_ser', 'cycles', lambda x: Protolog._median_cycles(x, 'ser')],
['cycles_des', 'cycles', lambda x: Protolog._median_cycles(x, 'des')],
['cycles_enc', 'cycles', lambda x: Protolog._median_cycles(x, 'enc')],
['cycles_dec', 'cycles', lambda x: Protolog._median_cycles(x, 'dec')],
#['cycles_ser_arr', 'cycles', lambda x: np.array(x['ser'][1:]) - np.mean(x['nop'][1:])],
#['cycles_des_arr', 'cycles', lambda x: np.array(x['des'][1:]) - np.mean(x['nop'][1:])],
#['cycles_enc_arr', 'cycles', lambda x: np.array(x['enc'][1:]) - np.mean(x['nop'][1:])],
#['cycles_dec_arr', 'cycles', lambda x: np.array(x['dec'][1:]) - np.mean(x['nop'][1:])],
['data_nop', 'data_size_nop', idem],
['data_ser', 'data_size_ser', idem],
['data_serdes', 'data_size_serdes', idem],
['heap_ser', 'heap_usage_ser', idem],
['heap_des', 'heap_usage_des', idem],
['serialized_size', 'serialized_size', idem],
['stack_alloc_ser', 'stack_online_ser', lambda x: x['allocated']],
['stack_set_ser', 'stack_online_ser', lambda x: x['used']],
['stack_alloc_des', 'stack_online_des', lambda x: x['allocated']],
['stack_set_des', 'stack_online_des', lambda x: x['used']],
['text_nop', 'text_size_nop', idem],
['text_ser', 'text_size_ser', idem],
['text_serdes', 'text_size_serdes', idem],
]
def __init__(self, logfile):
"""
Load and enrich raw protobench log data.
The enriched data can be accessed via the .aggregate class member,
see the class documentation for details.
"""
with open(logfile, 'rb') as f:
self.data = ubjson.load(f)
self.libraries = set()
self.architectures = set()
self.aggregate = dict()
for arch_lib in self.data.keys():
arch, lib, libopts = arch_lib.split(':')
library = lib + ':' + libopts
for benchmark in self.data[arch_lib].keys():
for benchmark_item in self.data[arch_lib][benchmark].keys():
subv = self.data[arch_lib][benchmark][benchmark_item]
for aggregate_label, data_label, getter in Protolog.datamap:
try:
self.add_datapoint(arch, library, (benchmark, benchmark_item), subv, aggregate_label, data_label, getter)
except KeyError:
pass
except TypeError as e:
print('TypeError in {} {} {} {}: {} -> {}'.format(
arch_lib, benchmark, benchmark_item, aggregate_label,
subv[data_label]['v'], str(e)))
pass
codegen = codegen_for_lib(lib, libopts.split(','), subv['data'])
if codegen.max_serialized_bytes != None:
self.add_datapoint(arch, library, (benchmark, benchmark_item), subv, 'buffer_size', data_label, lambda x: codegen.max_serialized_bytes)
else:
self.add_datapoint(arch, library, (benchmark, benchmark_item), subv, 'buffer_size', data_label, lambda x: 0)
#self.aggregate[(benchmark, benchmark_item)][arch][lib][aggregate_label] = getter(value[data_label]['v'])
for key in self.aggregate.keys():
for arch in self.aggregate[key].keys():
for lib, val in self.aggregate[key][arch].items():
try:
val['cycles_encser'] = val['cycles_enc'] + val['cycles_ser']
except KeyError:
pass
try:
val['cycles_desdec'] = val['cycles_des'] + val['cycles_dec']
except KeyError:
pass
try:
val['total_dmem_ser'] = val['stack_alloc_ser']
val['total_dmem_ser'] += val['heap_ser']
except KeyError:
pass
try:
val['total_dmem_des'] = val['stack_alloc_des']
val['total_dmem_des'] += val['heap_des']
except KeyError:
pass
try:
val['total_dmem_serdes'] = max(val['total_dmem_ser'], val['total_dmem_des'])
except KeyError:
pass
try:
val['text_ser_delta'] = val['text_ser'] - val['text_nop']
val['text_serdes_delta'] = val['text_serdes'] - val['text_nop']
except KeyError:
pass
try:
val['bss_ser_delta'] = val['bss_ser'] - val['bss_nop']
val['bss_serdes_delta'] = val['bss_serdes'] - val['bss_nop']
except KeyError:
pass
try:
val['data_ser_delta'] = val['data_ser'] - val['data_nop']
val['data_serdes_delta'] = val['data_serdes'] - val['data_nop']
except KeyError:
pass
try:
val['allmem_ser'] = val['text_ser'] + val['data_ser'] + val['bss_ser'] + val['total_dmem_ser'] - val['buffer_size']
val['allmem_serdes'] = val['text_serdes'] + val['data_serdes'] + val['bss_serdes'] + val['total_dmem_serdes'] - val['buffer_size']
except KeyError:
pass
def add_datapoint(self, arch, lib, key, value, aggregate_label, data_label, getter):
"""
Set self.aggregate[key][arch][lib][aggregate_Label] = getter(value[data_label]['v']).
Additionally, add lib to self.libraries and arch to self.architectures
key usually is ('benchmark name', 'sub-benchmark index').
"""
if data_label in value and 'v' in value[data_label]:
self.architectures.add(arch)
self.libraries.add(lib)
if not key in self.aggregate:
self.aggregate[key] = dict()
if not arch in self.aggregate[key]:
self.aggregate[key][arch] = dict()
if not lib in self.aggregate[key][arch]:
self.aggregate[key][arch][lib] = dict()
self.aggregate[key][arch][lib][aggregate_label] = getter(value[data_label]['v'])
|
