|
1
|
|
|
import os |
|
2
|
|
|
import logging |
|
3
|
|
|
import pdb |
|
4
|
|
|
import time |
|
5
|
|
|
import random |
|
6
|
|
|
import string |
|
7
|
|
|
import json |
|
8
|
|
|
import csv |
|
9
|
|
|
from multiprocessing import Process |
|
10
|
|
|
import numpy as np |
|
11
|
|
|
import concurrent.futures |
|
12
|
|
|
from client import MilvusClient |
|
13
|
|
|
import utils |
|
14
|
|
|
import parser |
|
15
|
|
|
from runner import Runner |
|
16
|
|
|
|
|
17
|
|
|
DELETE_INTERVAL_TIME = 5 |
|
18
|
|
|
INSERT_INTERVAL = 50000 |
|
19
|
|
|
logger = logging.getLogger("milvus_benchmark.local_runner") |
|
20
|
|
|
|
|
21
|
|
|
|
|
22
|
|
|
class LocalRunner(Runner): |
|
23
|
|
|
def __init__(self, host, port): |
|
24
|
|
|
""" |
|
25
|
|
|
Run tests at local mode. |
|
26
|
|
|
|
|
27
|
|
|
Make sure the server has started |
|
28
|
|
|
""" |
|
29
|
|
|
super(LocalRunner, self).__init__() |
|
30
|
|
|
self.host = host |
|
31
|
|
|
self.port = port |
|
32
|
|
|
|
|
33
|
|
|
def run(self, run_type, collection): |
|
34
|
|
|
logger.debug(run_type) |
|
35
|
|
|
logger.debug(collection) |
|
36
|
|
|
collection_name = collection["collection_name"] if "collection_name" in collection else None |
|
37
|
|
|
milvus_instance = MilvusClient(collection_name=collection_name, host=self.host, port=self.port) |
|
38
|
|
|
logger.info(milvus_instance.show_collections()) |
|
39
|
|
|
env_value = milvus_instance.get_server_config() |
|
40
|
|
|
logger.debug(env_value) |
|
41
|
|
|
|
|
42
|
|
|
if run_type in ["insert_performance", "insert_flush_performance"]: |
|
43
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
44
|
|
|
ni_per = collection["ni_per"] |
|
45
|
|
|
build_index = collection["build_index"] |
|
46
|
|
|
if milvus_instance.exists_collection(): |
|
47
|
|
|
milvus_instance.drop() |
|
48
|
|
|
time.sleep(10) |
|
49
|
|
|
milvus_instance.create_collection(collection_name, dimension, index_file_size, metric_type) |
|
50
|
|
|
if build_index is True: |
|
51
|
|
|
index_type = collection["index_type"] |
|
52
|
|
|
index_param = collection["index_param"] |
|
53
|
|
|
milvus_instance.create_index(index_type, index_param) |
|
54
|
|
|
logger.debug(milvus_instance.describe_index()) |
|
55
|
|
|
res = self.do_insert(milvus_instance, collection_name, data_type, dimension, collection_size, ni_per) |
|
56
|
|
|
milvus_instance.flush() |
|
57
|
|
|
logger.debug("Table row counts: %d" % milvus_instance.count()) |
|
58
|
|
|
if build_index is True: |
|
59
|
|
|
logger.debug("Start build index for last file") |
|
60
|
|
|
milvus_instance.create_index(index_type, index_param) |
|
|
|
|
|
|
61
|
|
|
logger.debug(milvus_instance.describe_index()) |
|
62
|
|
|
|
|
63
|
|
|
elif run_type == "delete_performance": |
|
64
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
65
|
|
|
ni_per = collection["ni_per"] |
|
66
|
|
|
if not milvus_instance.exists_collection(): |
|
67
|
|
|
logger.error(milvus_instance.show_collections()) |
|
68
|
|
|
logger.warning("Table: %s not found" % collection_name) |
|
69
|
|
|
return |
|
70
|
|
|
length = milvus_instance.count() |
|
71
|
|
|
ids = [i for i in range(length)] |
|
72
|
|
|
loops = int(length / ni_per) |
|
73
|
|
|
for i in range(loops): |
|
74
|
|
|
delete_ids = ids[i*ni_per : i*ni_per+ni_per] |
|
75
|
|
|
logger.debug("Delete %d - %d" % (delete_ids[0], delete_ids[-1])) |
|
76
|
|
|
milvus_instance.delete(delete_ids) |
|
77
|
|
|
milvus_instance.flush() |
|
78
|
|
|
logger.debug("Table row counts: %d" % milvus_instance.count()) |
|
79
|
|
|
logger.debug("Table row counts: %d" % milvus_instance.count()) |
|
80
|
|
|
milvus_instance.flush() |
|
81
|
|
|
logger.debug("Table row counts: %d" % milvus_instance.count()) |
|
82
|
|
|
|
|
83
|
|
|
elif run_type == "build_performance": |
|
84
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
85
|
|
|
index_type = collection["index_type"] |
|
86
|
|
|
index_param = collection["index_param"] |
|
87
|
|
|
if not milvus_instance.exists_collection(): |
|
88
|
|
|
logger.error("Table name: %s not existed" % collection_name) |
|
89
|
|
|
return |
|
90
|
|
|
search_params = {} |
|
91
|
|
|
start_time = time.time() |
|
92
|
|
|
# drop index |
|
93
|
|
|
logger.debug("Drop index") |
|
94
|
|
|
milvus_instance.drop_index() |
|
95
|
|
|
start_mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
96
|
|
|
milvus_instance.create_index(index_type, index_param) |
|
97
|
|
|
logger.debug(milvus_instance.describe_index()) |
|
98
|
|
|
logger.debug("Table row counts: %d" % milvus_instance.count()) |
|
99
|
|
|
end_time = time.time() |
|
100
|
|
|
end_mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
101
|
|
|
logger.debug("Diff memory: %s, current memory usage: %s, build time: %s" % ((end_mem_usage - start_mem_usage), end_mem_usage, round(end_time - start_time, 1))) |
|
102
|
|
|
|
|
103
|
|
|
elif run_type == "search_performance": |
|
104
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
105
|
|
|
run_count = collection["run_count"] |
|
106
|
|
|
top_ks = collection["top_ks"] |
|
107
|
|
|
nqs = collection["nqs"] |
|
108
|
|
|
search_params = collection["search_params"] |
|
109
|
|
|
# for debugging |
|
110
|
|
|
# time.sleep(3600) |
|
111
|
|
|
if not milvus_instance.exists_collection(): |
|
112
|
|
|
logger.error("Table name: %s not existed" % collection_name) |
|
113
|
|
|
return |
|
114
|
|
|
logger.info(milvus_instance.count()) |
|
115
|
|
|
result = milvus_instance.describe_index() |
|
116
|
|
|
logger.info(result) |
|
117
|
|
|
milvus_instance.preload_collection() |
|
118
|
|
|
mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
119
|
|
|
logger.info(mem_usage) |
|
120
|
|
|
for search_param in search_params: |
|
121
|
|
|
logger.info("Search param: %s" % json.dumps(search_param)) |
|
122
|
|
|
res = self.do_query(milvus_instance, collection_name, top_ks, nqs, run_count, search_param) |
|
123
|
|
|
headers = ["Nq/Top-k"] |
|
124
|
|
|
headers.extend([str(top_k) for top_k in top_ks]) |
|
125
|
|
|
logger.info("Search param: %s" % json.dumps(search_param)) |
|
126
|
|
|
utils.print_table(headers, nqs, res) |
|
127
|
|
|
mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
128
|
|
|
logger.info(mem_usage) |
|
129
|
|
|
|
|
130
|
|
|
elif run_type == "locust_search_performance": |
|
131
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
132
|
|
|
### spawn locust requests |
|
133
|
|
|
collection_num = collection["collection_num"] |
|
134
|
|
|
task = collection["task"] |
|
135
|
|
|
#. generate task code |
|
136
|
|
|
task_file = utils.get_unique_name() |
|
137
|
|
|
task_file_script = task_file+'.py' |
|
138
|
|
|
task_file_csv = task_file+'_stats.csv' |
|
139
|
|
|
task_type = task["type"] |
|
140
|
|
|
connection_type = "single" |
|
141
|
|
|
connection_num = task["connection_num"] |
|
142
|
|
|
if connection_num > 1: |
|
143
|
|
|
connection_type = "multi" |
|
144
|
|
|
clients_num = task["clients_num"] |
|
145
|
|
|
hatch_rate = task["hatch_rate"] |
|
146
|
|
|
during_time = task["during_time"] |
|
147
|
|
|
def_name = task_type |
|
148
|
|
|
task_params = task["params"] |
|
149
|
|
|
collection_names = [] |
|
150
|
|
|
for i in range(collection_num): |
|
151
|
|
|
suffix = "".join(random.choice(string.ascii_letters + string.digits) for _ in range(5)) |
|
152
|
|
|
collection_names.append(collection_name + "_" + suffix) |
|
153
|
|
|
# collection_names = ['sift_1m_1024_128_l2_Kg6co', 'sift_1m_1024_128_l2_egkBK', 'sift_1m_1024_128_l2_D0wtE', |
|
154
|
|
|
# 'sift_1m_1024_128_l2_9naps', 'sift_1m_1024_128_l2_iJ0jj', 'sift_1m_1024_128_l2_nqUTm', |
|
155
|
|
|
# 'sift_1m_1024_128_l2_GIF0D', 'sift_1m_1024_128_l2_EL2qk', 'sift_1m_1024_128_l2_qLRnC', |
|
156
|
|
|
# 'sift_1m_1024_128_l2_8Ditg'] |
|
157
|
|
|
# ##### |
|
158
|
|
|
ni_per = collection["ni_per"] |
|
159
|
|
|
build_index = collection["build_index"] |
|
160
|
|
|
for c_name in collection_names: |
|
161
|
|
|
milvus_instance = MilvusClient(collection_name=c_name, host=self.host, port=self.port) |
|
162
|
|
|
if milvus_instance.exists_collection(collection_name=c_name): |
|
163
|
|
|
milvus_instance.drop(name=c_name) |
|
164
|
|
|
time.sleep(10) |
|
165
|
|
|
milvus_instance.create_collection(c_name, dimension, index_file_size, metric_type) |
|
166
|
|
|
if build_index is True: |
|
167
|
|
|
index_type = collection["index_type"] |
|
168
|
|
|
index_param = collection["index_param"] |
|
169
|
|
|
milvus_instance.create_index(index_type, index_param) |
|
170
|
|
|
logger.debug(milvus_instance.describe_index()) |
|
171
|
|
|
res = self.do_insert(milvus_instance, c_name, data_type, dimension, collection_size, ni_per) |
|
172
|
|
|
milvus_instance.flush() |
|
173
|
|
|
logger.debug("Table row counts: %d" % milvus_instance.count(name=c_name)) |
|
174
|
|
|
if build_index is True: |
|
175
|
|
|
logger.debug("Start build index for last file") |
|
176
|
|
|
milvus_instance.create_index(index_type, index_param) |
|
177
|
|
|
logger.debug(milvus_instance.describe_index()) |
|
178
|
|
|
code_str = """ |
|
179
|
|
|
import random |
|
180
|
|
|
import string |
|
181
|
|
|
from locust import User, task, between |
|
182
|
|
|
from locust_task import MilvusTask |
|
183
|
|
|
from client import MilvusClient |
|
184
|
|
|
|
|
185
|
|
|
host = '%s' |
|
186
|
|
|
port = %s |
|
187
|
|
|
dim = %s |
|
188
|
|
|
connection_type = '%s' |
|
189
|
|
|
collection_names = %s |
|
190
|
|
|
m = MilvusClient(host=host, port=port) |
|
191
|
|
|
|
|
192
|
|
|
|
|
193
|
|
|
def get_collection_name(): |
|
194
|
|
|
return random.choice(collection_names) |
|
195
|
|
|
|
|
196
|
|
|
|
|
197
|
|
|
def get_client(collection_name): |
|
198
|
|
|
if connection_type == 'single': |
|
199
|
|
|
return MilvusTask(m=m) |
|
200
|
|
|
elif connection_type == 'multi': |
|
201
|
|
|
return MilvusTask(connection_type='multi', host=host, port=port, collection_name=collection_name) |
|
202
|
|
|
|
|
203
|
|
|
|
|
204
|
|
|
class QueryTask(User): |
|
205
|
|
|
wait_time = between(0.001, 0.002) |
|
206
|
|
|
|
|
207
|
|
|
@task() |
|
208
|
|
|
def %s(self): |
|
209
|
|
|
top_k = %s |
|
210
|
|
|
X = [[random.random() for i in range(dim)] for i in range(%s)] |
|
211
|
|
|
search_param = %s |
|
212
|
|
|
collection_name = get_collection_name() |
|
213
|
|
|
print(collection_name) |
|
214
|
|
|
client = get_client(collection_name) |
|
215
|
|
|
client.query(X, top_k, search_param, collection_name=collection_name) |
|
216
|
|
|
""" % (self.host, self.port, dimension, connection_type, collection_names, def_name, task_params["top_k"], task_params["nq"], task_params["search_param"]) |
|
217
|
|
|
with open(task_file_script, 'w+') as fd: |
|
218
|
|
|
fd.write(code_str) |
|
219
|
|
|
locust_cmd = "locust -f %s --headless --csv=%s -u %d -r %d -t %s" % ( |
|
220
|
|
|
task_file_script, |
|
221
|
|
|
task_file, |
|
222
|
|
|
clients_num, |
|
223
|
|
|
hatch_rate, |
|
224
|
|
|
during_time) |
|
225
|
|
|
logger.info(locust_cmd) |
|
226
|
|
|
try: |
|
227
|
|
|
res = os.system(locust_cmd) |
|
228
|
|
|
except Exception as e: |
|
229
|
|
|
logger.error(str(e)) |
|
230
|
|
|
return |
|
231
|
|
|
#. retrieve and collect test statistics |
|
232
|
|
|
metric = None |
|
233
|
|
|
with open(task_file_csv, newline='') as fd: |
|
234
|
|
|
dr = csv.DictReader(fd) |
|
235
|
|
|
for row in dr: |
|
236
|
|
|
if row["Name"] != "Aggregated": |
|
237
|
|
|
continue |
|
238
|
|
|
metric = row |
|
239
|
|
|
logger.info(metric) |
|
240
|
|
|
# clean up temp files |
|
241
|
|
|
|
|
242
|
|
|
elif run_type == "search_ids_stability": |
|
243
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
244
|
|
|
search_params = collection["search_params"] |
|
245
|
|
|
during_time = collection["during_time"] |
|
246
|
|
|
ids_length = collection["ids_length"] |
|
247
|
|
|
ids = collection["ids"] |
|
248
|
|
|
logger.info(milvus_instance.count()) |
|
249
|
|
|
index_info = milvus_instance.describe_index() |
|
250
|
|
|
logger.info(index_info) |
|
251
|
|
|
g_top_k = int(collection["top_ks"].split("-")[1]) |
|
252
|
|
|
l_top_k = int(collection["top_ks"].split("-")[0]) |
|
253
|
|
|
g_id = int(ids.split("-")[1]) |
|
254
|
|
|
l_id = int(ids.split("-")[0]) |
|
255
|
|
|
g_id_length = int(ids_length.split("-")[1]) |
|
256
|
|
|
l_id_length = int(ids_length.split("-")[0]) |
|
257
|
|
|
|
|
258
|
|
|
milvus_instance.preload_collection() |
|
259
|
|
|
start_mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
260
|
|
|
logger.debug(start_mem_usage) |
|
261
|
|
|
start_time = time.time() |
|
262
|
|
|
while time.time() < start_time + during_time * 60: |
|
263
|
|
|
search_param = {} |
|
264
|
|
|
top_k = random.randint(l_top_k, g_top_k) |
|
265
|
|
|
ids_num = random.randint(l_id_length, g_id_length) |
|
266
|
|
|
l_ids = random.randint(l_id, g_id-ids_num) |
|
267
|
|
|
# ids_param = [random.randint(l_id_length, g_id_length) for _ in range(ids_num)] |
|
268
|
|
|
ids_param = [id for id in range(l_ids, l_ids+ids_num)] |
|
269
|
|
|
for k, v in search_params.items(): |
|
270
|
|
|
search_param[k] = random.randint(int(v.split("-")[0]), int(v.split("-")[1])) |
|
271
|
|
|
logger.debug("Query top-k: %d, ids_num: %d, param: %s" % (top_k, ids_num, json.dumps(search_param))) |
|
272
|
|
|
result = milvus_instance.query_ids(top_k, ids_param, search_param=search_param) |
|
273
|
|
|
end_mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
274
|
|
|
metrics = { |
|
275
|
|
|
"during_time": during_time, |
|
276
|
|
|
"start_mem_usage": start_mem_usage, |
|
277
|
|
|
"end_mem_usage": end_mem_usage, |
|
278
|
|
|
"diff_mem": end_mem_usage - start_mem_usage, |
|
279
|
|
|
} |
|
280
|
|
|
logger.info(metrics) |
|
281
|
|
|
|
|
282
|
|
|
elif run_type == "search_performance_concurrents": |
|
283
|
|
|
data_type, dimension, metric_type = parser.parse_ann_collection_name(collection_name) |
|
284
|
|
|
hdf5_source_file = collection["source_file"] |
|
285
|
|
|
use_single_connection = collection["use_single_connection"] |
|
286
|
|
|
concurrents = collection["concurrents"] |
|
287
|
|
|
top_ks = collection["top_ks"] |
|
288
|
|
|
nqs = collection["nqs"] |
|
289
|
|
|
search_params = self.generate_combinations(collection["search_params"]) |
|
290
|
|
|
if not milvus_instance.exists_collection(): |
|
291
|
|
|
logger.error("Table name: %s not existed" % collection_name) |
|
292
|
|
|
return |
|
293
|
|
|
logger.info(milvus_instance.count()) |
|
294
|
|
|
result = milvus_instance.describe_index() |
|
295
|
|
|
logger.info(result) |
|
296
|
|
|
milvus_instance.preload_collection() |
|
297
|
|
|
dataset = utils.get_dataset(hdf5_source_file) |
|
298
|
|
|
for concurrent_num in concurrents: |
|
299
|
|
|
top_k = top_ks[0] |
|
300
|
|
|
for nq in nqs: |
|
301
|
|
|
mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
302
|
|
|
logger.info(mem_usage) |
|
303
|
|
|
query_vectors = self.normalize(metric_type, np.array(dataset["test"][:nq])) |
|
304
|
|
|
logger.debug(search_params) |
|
305
|
|
|
for search_param in search_params: |
|
306
|
|
|
logger.info("Search param: %s" % json.dumps(search_param)) |
|
307
|
|
|
total_time = 0.0 |
|
308
|
|
|
if use_single_connection is True: |
|
309
|
|
|
connections = [MilvusClient(collection_name=collection_name, host=self.host, port=self.port)] |
|
310
|
|
|
with concurrent.futures.ThreadPoolExecutor(max_workers=concurrent_num) as executor: |
|
311
|
|
|
future_results = {executor.submit( |
|
312
|
|
|
self.do_query_qps, connections[0], query_vectors, top_k, search_param=search_param) : index for index in range(concurrent_num)} |
|
313
|
|
|
else: |
|
314
|
|
|
connections = [MilvusClient(collection_name=collection_name, host=self.hos, port=self.port) for i in range(concurrent_num)] |
|
315
|
|
|
with concurrent.futures.ThreadPoolExecutor(max_workers=concurrent_num) as executor: |
|
316
|
|
|
future_results = {executor.submit( |
|
317
|
|
|
self.do_query_qps, connections[index], query_vectors, top_k, search_param=search_param) : index for index in range(concurrent_num)} |
|
318
|
|
|
for future in concurrent.futures.as_completed(future_results): |
|
319
|
|
|
total_time = total_time + future.result() |
|
320
|
|
|
qps_value = total_time / concurrent_num |
|
321
|
|
|
logger.debug("QPS value: %f, total_time: %f, request_nums: %f" % (qps_value, total_time, concurrent_num)) |
|
322
|
|
|
mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
323
|
|
|
logger.info(mem_usage) |
|
324
|
|
|
|
|
325
|
|
|
elif run_type == "ann_accuracy": |
|
326
|
|
|
hdf5_source_file = collection["source_file"] |
|
327
|
|
|
collection_name = collection["collection_name"] |
|
328
|
|
|
index_file_sizes = collection["index_file_sizes"] |
|
329
|
|
|
index_types = collection["index_types"] |
|
330
|
|
|
index_params = collection["index_params"] |
|
331
|
|
|
top_ks = collection["top_ks"] |
|
332
|
|
|
nqs = collection["nqs"] |
|
333
|
|
|
search_params = collection["search_params"] |
|
334
|
|
|
# mapping to search param list |
|
335
|
|
|
search_params = self.generate_combinations(search_params) |
|
336
|
|
|
# mapping to index param list |
|
337
|
|
|
index_params = self.generate_combinations(index_params) |
|
338
|
|
|
|
|
339
|
|
|
data_type, dimension, metric_type = parser.parse_ann_collection_name(collection_name) |
|
340
|
|
|
dataset = utils.get_dataset(hdf5_source_file) |
|
341
|
|
|
if milvus_instance.exists_collection(collection_name): |
|
342
|
|
|
logger.info("Re-create collection: %s" % collection_name) |
|
343
|
|
|
milvus_instance.drop() |
|
344
|
|
|
time.sleep(DELETE_INTERVAL_TIME) |
|
345
|
|
|
true_ids = np.array(dataset["neighbors"]) |
|
346
|
|
|
for index_file_size in index_file_sizes: |
|
347
|
|
|
milvus_instance.create_collection(collection_name, dimension, index_file_size, metric_type) |
|
348
|
|
|
logger.info(milvus_instance.describe()) |
|
349
|
|
|
insert_vectors = self.normalize(metric_type, np.array(dataset["train"])) |
|
350
|
|
|
logger.debug(len(insert_vectors)) |
|
351
|
|
|
# Insert batch once |
|
352
|
|
|
# milvus_instance.insert(insert_vectors) |
|
353
|
|
|
loops = len(insert_vectors) // INSERT_INTERVAL + 1 |
|
354
|
|
|
for i in range(loops): |
|
355
|
|
|
start = i*INSERT_INTERVAL |
|
356
|
|
|
end = min((i+1)*INSERT_INTERVAL, len(insert_vectors)) |
|
357
|
|
|
tmp_vectors = insert_vectors[start:end] |
|
358
|
|
|
if start < end: |
|
359
|
|
|
if not isinstance(tmp_vectors, list): |
|
360
|
|
|
milvus_instance.insert(tmp_vectors.tolist(), ids=[i for i in range(start, end)]) |
|
361
|
|
|
else: |
|
362
|
|
|
milvus_instance.insert(tmp_vectors, ids=[i for i in range(start, end)]) |
|
363
|
|
|
milvus_instance.flush() |
|
364
|
|
|
logger.info("Table: %s, row count: %s" % (collection_name, milvus_instance.count())) |
|
365
|
|
|
if milvus_instance.count() != len(insert_vectors): |
|
366
|
|
|
logger.error("Table row count is not equal to insert vectors") |
|
367
|
|
|
return |
|
368
|
|
|
for index_type in index_types: |
|
369
|
|
|
for index_param in index_params: |
|
370
|
|
|
logger.debug("Building index with param: %s" % json.dumps(index_param)) |
|
371
|
|
|
milvus_instance.create_index(index_type, index_param=index_param) |
|
372
|
|
|
logger.info(milvus_instance.describe_index()) |
|
373
|
|
|
logger.info("Start preload collection: %s" % collection_name) |
|
374
|
|
|
milvus_instance.preload_collection() |
|
375
|
|
|
for search_param in search_params: |
|
376
|
|
|
for nq in nqs: |
|
377
|
|
|
query_vectors = self.normalize(metric_type, np.array(dataset["test"][:nq])) |
|
378
|
|
|
for top_k in top_ks: |
|
379
|
|
|
logger.debug("Search nq: %d, top-k: %d, search_param: %s" % (nq, top_k, json.dumps(search_param))) |
|
380
|
|
|
if not isinstance(query_vectors, list): |
|
381
|
|
|
result = milvus_instance.query(query_vectors.tolist(), top_k, search_param=search_param) |
|
382
|
|
|
else: |
|
383
|
|
|
result = milvus_instance.query(query_vectors, top_k, search_param=search_param) |
|
384
|
|
|
result_ids = result.id_array |
|
385
|
|
|
acc_value = self.get_recall_value(true_ids[:nq, :top_k].tolist(), result_ids) |
|
386
|
|
|
logger.info("Query ann_accuracy: %s" % acc_value) |
|
387
|
|
|
|
|
388
|
|
|
|
|
389
|
|
|
elif run_type == "stability": |
|
390
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
391
|
|
|
search_params = collection["search_params"] |
|
392
|
|
|
insert_xb = collection["insert_xb"] |
|
393
|
|
|
insert_interval = collection["insert_interval"] |
|
394
|
|
|
delete_xb = collection["delete_xb"] |
|
395
|
|
|
# flush_interval = collection["flush_interval"] |
|
396
|
|
|
# compact_interval = collection["compact_interval"] |
|
397
|
|
|
during_time = collection["during_time"] |
|
398
|
|
|
if not milvus_instance.exists_collection(): |
|
399
|
|
|
logger.error(milvus_instance.show_collections()) |
|
400
|
|
|
logger.error("Table name: %s not existed" % collection_name) |
|
401
|
|
|
return |
|
402
|
|
|
g_top_k = int(collection["top_ks"].split("-")[1]) |
|
403
|
|
|
g_nq = int(collection["nqs"].split("-")[1]) |
|
404
|
|
|
l_top_k = int(collection["top_ks"].split("-")[0]) |
|
405
|
|
|
l_nq = int(collection["nqs"].split("-")[0]) |
|
406
|
|
|
milvus_instance.preload_collection() |
|
407
|
|
|
start_mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
408
|
|
|
start_row_count = milvus_instance.count() |
|
409
|
|
|
logger.debug(milvus_instance.describe_index()) |
|
410
|
|
|
logger.info(start_row_count) |
|
411
|
|
|
start_time = time.time() |
|
412
|
|
|
i = 0 |
|
413
|
|
|
ids = [] |
|
414
|
|
|
insert_vectors = [[random.random() for _ in range(dimension)] for _ in range(insert_xb)] |
|
415
|
|
|
query_vectors = [[random.random() for _ in range(dimension)] for _ in range(10000)] |
|
416
|
|
View Code Duplication |
while time.time() < start_time + during_time * 60: |
|
|
|
|
|
|
417
|
|
|
i = i + 1 |
|
418
|
|
|
for _ in range(insert_interval): |
|
419
|
|
|
top_k = random.randint(l_top_k, g_top_k) |
|
420
|
|
|
nq = random.randint(l_nq, g_nq) |
|
421
|
|
|
search_param = {} |
|
422
|
|
|
for k, v in search_params.items(): |
|
423
|
|
|
search_param[k] = random.randint(int(v.split("-")[0]), int(v.split("-")[1])) |
|
424
|
|
|
logger.debug("Query nq: %d, top-k: %d, param: %s" % (nq, top_k, json.dumps(search_param))) |
|
425
|
|
|
result = milvus_instance.query(query_vectors[0:nq], top_k, search_param=search_param) |
|
426
|
|
|
count = milvus_instance.count() |
|
427
|
|
|
insert_ids = [(count+x) for x in range(len(insert_vectors))] |
|
428
|
|
|
ids.extend(insert_ids) |
|
429
|
|
|
_, res = milvus_instance.insert(insert_vectors, ids=insert_ids) |
|
430
|
|
|
logger.debug("%d, row_count: %d" % (i, milvus_instance.count())) |
|
431
|
|
|
milvus_instance.delete(ids[-delete_xb:]) |
|
432
|
|
|
milvus_instance.flush() |
|
433
|
|
|
milvus_instance.compact() |
|
434
|
|
|
end_mem_usage = milvus_instance.get_mem_info()["memory_used"] |
|
435
|
|
|
end_row_count = milvus_instance.count() |
|
436
|
|
|
metrics = { |
|
437
|
|
|
"during_time": during_time, |
|
438
|
|
|
"start_mem_usage": start_mem_usage, |
|
439
|
|
|
"end_mem_usage": end_mem_usage, |
|
440
|
|
|
"diff_mem": end_mem_usage - start_mem_usage, |
|
441
|
|
|
"row_count_increments": end_row_count - start_row_count |
|
442
|
|
|
} |
|
443
|
|
|
logger.info(metrics) |
|
444
|
|
|
|
|
445
|
|
|
elif run_type == "loop_stability": |
|
446
|
|
|
# init data |
|
447
|
|
|
milvus_instance.clean_db() |
|
448
|
|
|
pull_interval = collection["pull_interval"] |
|
449
|
|
|
pull_interval_seconds = pull_interval * 60 |
|
450
|
|
|
collection_num = collection["collection_num"] |
|
451
|
|
|
dimension = collection["dimension"] if "dimension" in collection else 128 |
|
452
|
|
|
insert_xb = collection["insert_xb"] if "insert_xb" in collection else 100000 |
|
453
|
|
|
index_types = collection["index_types"] if "index_types" in collection else ['ivf_sq8'] |
|
454
|
|
|
index_param = {"nlist": 2048} |
|
455
|
|
|
collection_names = [] |
|
456
|
|
|
milvus_instances_map = {} |
|
457
|
|
|
insert_vectors = [[random.random() for _ in range(dimension)] for _ in range(insert_xb)] |
|
458
|
|
|
for i in range(collection_num): |
|
459
|
|
|
name = utils.get_unique_name(prefix="collection_") |
|
460
|
|
|
collection_names.append(name) |
|
461
|
|
|
metric_type = random.choice(["l2", "ip"]) |
|
462
|
|
|
index_file_size = random.randint(10, 20) |
|
463
|
|
|
milvus_instance.create_collection(name, dimension, index_file_size, metric_type) |
|
464
|
|
|
milvus_instance = MilvusClient(collection_name=name, host=self.host) |
|
465
|
|
|
index_type = random.choice(index_types) |
|
466
|
|
|
milvus_instance.create_index(index_type, index_param=index_param) |
|
467
|
|
|
logger.info(milvus_instance.describe_index()) |
|
468
|
|
|
insert_vectors = utils.normalize(metric_type, insert_vectors) |
|
469
|
|
|
milvus_instance.insert(insert_vectors) |
|
470
|
|
|
milvus_instance.flush() |
|
471
|
|
|
milvus_instances_map.update({name: milvus_instance}) |
|
472
|
|
|
logger.info(milvus_instance.describe_index()) |
|
473
|
|
|
logger.info(milvus_instance.describe()) |
|
474
|
|
|
|
|
475
|
|
|
tasks = ["insert_rand", "delete_rand", "query_rand", "flush"] |
|
476
|
|
|
i = 1 |
|
477
|
|
|
while True: |
|
478
|
|
|
logger.info("Loop time: %d" % i) |
|
479
|
|
|
start_time = time.time() |
|
480
|
|
|
while time.time() - start_time < pull_interval_seconds: |
|
481
|
|
|
# choose collection |
|
482
|
|
|
tmp_collection_name = random.choice(collection_names) |
|
483
|
|
|
# choose task from task |
|
484
|
|
|
task_name = random.choice(tasks) |
|
485
|
|
|
logger.info(tmp_collection_name) |
|
486
|
|
|
logger.info(task_name) |
|
487
|
|
|
# execute task |
|
488
|
|
|
task_run = getattr(milvus_instances_map[tmp_collection_name], task_name) |
|
489
|
|
|
task_run() |
|
490
|
|
|
# new connection |
|
491
|
|
|
for name in collection_names: |
|
492
|
|
|
milvus_instance = MilvusClient(collection_name=name, host=self.host) |
|
493
|
|
|
milvus_instances_map.update({name: milvus_instance}) |
|
494
|
|
|
i = i + 1 |
|
495
|
|
|
elif run_type == "locust_mix_performance": |
|
496
|
|
|
(data_type, collection_size, index_file_size, dimension, metric_type) = parser.collection_parser(collection_name) |
|
497
|
|
|
# ni_per = collection["ni_per"] |
|
498
|
|
|
# build_index = collection["build_index"] |
|
499
|
|
|
# if milvus_instance.exists_collection(): |
|
500
|
|
|
# milvus_instance.drop() |
|
501
|
|
|
# time.sleep(10) |
|
502
|
|
|
# milvus_instance.create_collection(collection_name, dimension, index_file_size, metric_type) |
|
503
|
|
|
# if build_index is True: |
|
504
|
|
|
# index_type = collection["index_type"] |
|
505
|
|
|
# index_param = collection["index_param"] |
|
506
|
|
|
# milvus_instance.create_index(index_type, index_param) |
|
507
|
|
|
# logger.debug(milvus_instance.describe_index()) |
|
508
|
|
|
# res = self.do_insert(milvus_instance, collection_name, data_type, dimension, collection_size, ni_per) |
|
509
|
|
|
# milvus_instance.flush() |
|
510
|
|
|
# logger.debug("Table row counts: %d" % milvus_instance.count()) |
|
511
|
|
|
# if build_index is True: |
|
512
|
|
|
# logger.debug("Start build index for last file") |
|
513
|
|
|
# milvus_instance.create_index(index_type, index_param) |
|
514
|
|
|
# logger.debug(milvus_instance.describe_index()) |
|
515
|
|
|
task = collection["tasks"] |
|
516
|
|
|
task_file = utils.get_unique_name() |
|
517
|
|
|
task_file_script = task_file + '.py' |
|
518
|
|
|
task_file_csv = task_file + '_stats.csv' |
|
519
|
|
|
task_types = task["types"] |
|
520
|
|
|
connection_type = "single" |
|
521
|
|
|
connection_num = task["connection_num"] |
|
522
|
|
|
if connection_num > 1: |
|
523
|
|
|
connection_type = "multi" |
|
524
|
|
|
clients_num = task["clients_num"] |
|
525
|
|
|
hatch_rate = task["hatch_rate"] |
|
526
|
|
|
during_time = task["during_time"] |
|
527
|
|
|
def_strs = "" |
|
528
|
|
|
for task_type in task_types: |
|
529
|
|
|
_type = task_type["type"] |
|
530
|
|
|
weight = task_type["weight"] |
|
531
|
|
|
if _type == "flush": |
|
532
|
|
|
def_str = """ |
|
533
|
|
|
@task(%d) |
|
534
|
|
|
def flush(self): |
|
535
|
|
|
client = get_client(collection_name) |
|
536
|
|
|
client.flush(collection_name=collection_name) |
|
537
|
|
|
""" % weight |
|
538
|
|
|
if _type == "compact": |
|
539
|
|
|
def_str = """ |
|
540
|
|
|
@task(%d) |
|
541
|
|
|
def compact(self): |
|
542
|
|
|
client = get_client(collection_name) |
|
543
|
|
|
client.compact(collection_name) |
|
544
|
|
|
""" % weight |
|
545
|
|
|
if _type == "query": |
|
546
|
|
|
def_str = """ |
|
547
|
|
|
@task(%d) |
|
548
|
|
|
def query(self): |
|
549
|
|
|
client = get_client(collection_name) |
|
550
|
|
|
params = %s |
|
551
|
|
|
X = [[random.random() for i in range(dim)] for i in range(params["nq"])] |
|
552
|
|
|
client.query(X, params["top_k"], params["search_param"], collection_name=collection_name) |
|
553
|
|
|
""" % (weight, task_type["params"]) |
|
554
|
|
|
if _type == "insert": |
|
555
|
|
|
def_str = """ |
|
556
|
|
|
@task(%d) |
|
557
|
|
|
def insert(self): |
|
558
|
|
|
client = get_client(collection_name) |
|
559
|
|
|
params = %s |
|
560
|
|
|
ids = [random.randint(10, 1000000) for i in range(params["nb"])] |
|
561
|
|
|
X = [[random.random() for i in range(dim)] for i in range(params["nb"])] |
|
562
|
|
|
client.insert(X,ids=ids, collection_name=collection_name) |
|
563
|
|
|
""" % (weight, task_type["params"]) |
|
564
|
|
|
if _type == "delete": |
|
565
|
|
|
def_str = """ |
|
566
|
|
|
@task(%d) |
|
567
|
|
|
def delete(self): |
|
568
|
|
|
client = get_client(collection_name) |
|
569
|
|
|
ids = [random.randint(1, 1000000) for i in range(1)] |
|
570
|
|
|
client.delete(ids, collection_name) |
|
571
|
|
|
""" % weight |
|
572
|
|
|
def_strs += def_str |
|
|
|
|
|
|
573
|
|
|
print(def_strs) |
|
574
|
|
|
code_str = """ |
|
575
|
|
|
import random |
|
576
|
|
|
import json |
|
577
|
|
|
from locust import User, task, between |
|
578
|
|
|
from locust_task import MilvusTask |
|
579
|
|
|
from client import MilvusClient |
|
580
|
|
|
|
|
581
|
|
|
host = '%s' |
|
582
|
|
|
port = %s |
|
583
|
|
|
collection_name = '%s' |
|
584
|
|
|
dim = %s |
|
585
|
|
|
connection_type = '%s' |
|
586
|
|
|
m = MilvusClient(host=host, port=port) |
|
587
|
|
|
|
|
588
|
|
|
|
|
589
|
|
|
def get_client(collection_name): |
|
590
|
|
|
if connection_type == 'single': |
|
591
|
|
|
return MilvusTask(m=m) |
|
592
|
|
|
elif connection_type == 'multi': |
|
593
|
|
|
return MilvusTask(connection_type='multi', host=host, port=port, collection_name=collection_name) |
|
594
|
|
|
|
|
595
|
|
|
|
|
596
|
|
|
class MixTask(User): |
|
597
|
|
|
wait_time = between(0.001, 0.002) |
|
598
|
|
|
%s |
|
599
|
|
|
""" % (self.host, self.port, collection_name, dimension, connection_type, def_strs) |
|
600
|
|
|
with open(task_file_script, "w+") as fd: |
|
601
|
|
|
fd.write(code_str) |
|
|
|
|
|
|
602
|
|
|
locust_cmd = "locust -f %s --headless --csv=%s -u %d -r %d -t %s" % ( |
|
603
|
|
|
task_file_script, |
|
604
|
|
|
task_file, |
|
605
|
|
|
clients_num, |
|
606
|
|
|
hatch_rate, |
|
607
|
|
|
during_time) |
|
608
|
|
|
logger.info(locust_cmd) |
|
609
|
|
|
try: |
|
610
|
|
|
res = os.system(locust_cmd) |
|
611
|
|
|
except Exception as e: |
|
612
|
|
|
logger.error(str(e)) |
|
613
|
|
|
return |
|
614
|
|
|
|
|
615
|
|
|
# . retrieve and collect test statistics |
|
616
|
|
|
metric = None |
|
617
|
|
|
with open(task_file_csv, newline='') as fd: |
|
618
|
|
|
dr = csv.DictReader(fd) |
|
619
|
|
|
for row in dr: |
|
620
|
|
|
if row["Name"] != "Aggregated": |
|
621
|
|
|
continue |
|
622
|
|
|
metric = row |
|
623
|
|
|
logger.info(metric) |
|
624
|
|
|
|
|
625
|
|
|
else: |
|
626
|
|
|
logger.warning("Run type not defined") |
|
627
|
|
|
return |
|
628
|
|
|
logger.debug("Test finished") |
|
629
|
|
|
|
milvus-io /
milvus
Loading history...
