├── config_example.ini ├── configs.py ├── draw.py ├── environment.py ├── generate_sysbench_insert_and_delete_query.py ├── get_res_data_from_file.py ├── get_sysbench_workload_from_slow_file.py ├── get_workload_from_file.py ├── knob-effect-test.py ├── main.py ├── model.py ├── pre-processing └── knob_rankiing.py ├── requirements.txt ├── run_job.py ├── run_workload.py ├── saved_model_weights-141threads ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights-2021-07-22-2 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights-2021-07-22 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights ├── actor_weights.h5 ├── critic_weights.h5 └── rewards.txt ├── saved_model_weights_2021-07-07 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights_2021-07-08 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights_2021-07-09 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights_2021_07_02_23_39_12 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights_2021_07_11 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights_2021_07_11_2 ├── actor_weights.h5 └── critic_weights.h5 ├── saved_model_weights_tmp ├── actor_weights.h5 └── critic_weights.h5 ├── scripts ├── output.res └── run_job.sh ├── sql2resource.py ├── training-data └── trainData_sql.txt ├── workload ├── workload_read_10000.txt └── workload_sysbench_readonly_table20_tablesize8000000.txt ├── workload_file_example.txt └── 配置文档.md /config_example.ini: -------------------------------------------------------------------------------- 1 | [predictor] 2 | # SELECT count(*) FROM INFORMATION_SCHEMA.INNODB_METRICS where status='enabled'; 3 | predictor_output_dim = 65 4 | # predictor_epoch 5 | predictor_epoch = 100 6 | 7 | 8 | [database_tune] 9 | # Host IP Address 10 | host = XX.XXX.XX.9 11 | # Host Port Number 12 | port = 3306 13 | # Database User Name 14 | user = root 15 | # Database Name 16 | database = XXX 17 | # Database Password 18 | password = XXXXXX 19 | # Limit for the requests (queries) in a workload' 20 | num_event = 1000 21 | # Read percentage 22 | p_r_range = 0.6 23 | # Update percentage 24 | p_u_index = 0.2 25 | # Insert percentage 26 | p_i = 0.1 27 | # Delete percentage 28 | p_d = 0.1 29 | # Iteration Number 30 | num_trial = 500 31 | # Maximum sample number cached in RL 32 | maxlen_mem = 2000 33 | # maxlen_predict_mem 34 | maxlen_predict_mem = 2000 35 | # learning_rate 36 | learning_rate = 1e-3 37 | # Sample threshold to train RL 38 | train_min_size = 3 39 | # Training finish if the accumulated score is over the value 40 | stopping_score = 90 41 | # Training finish if the throughput improvement percentage is over the value 42 | stopping_throughput_improvement_percentage = 0.12 43 | # Training Performance Comparision 44 | linelist = ['res_predict-1619930491','res_random-1619930491'] 45 | # [Latency, Throughput] 46 | performance_metric = ['Latency'] 47 | # workload_file, like workload_file_example.txt 48 | workload_file_path = /XX/CXX/XXX/sysbench_workload.txt 49 | # thread_num 50 | thread_num = 500 51 | # thread_num dependent on the max connections 1/true 0/false 52 | thread_num_auto = 1 53 | 54 | [knob_config] 55 | # you should increase the value of the table_open_cache variable if the numer of opened tables is large 56 | table_open_cache = {'type':'infer','min_value':1,'max_value':2000,'length':1} 57 | # The maximum permitted number of simultaneous client connections 58 | max_connections = {'type':'infer','min_value': 20,'max_value': 100000,'length':1} 59 | # The minimum size of the buffer that is used for plain index scans, range index scans, and joins that do not use indexes and thus perform full table scans. 60 | join_buffer_size = {'type':'infer','min_value': 128,'max_value': 134217728,'length':32} 61 | # The size of the buffer that is allocated when preloading indexes. 62 | preload_buffer_size = {'type':'infer','min_value': 1024,'max_value': 1073741824,'length':32} 63 | # Each session that must perform a sort allocates a buffer of this size. 64 | sort_buffer_size = {'type':'infer','min_value': 32768,'max_value': 134217728,'length':32} 65 | # The size of the cache to hold changes to the binary log during a transaction. 66 | binlog_cache_size = {'type':'infer','min_value': 4096,'max_value': 134217728,'length':32} 67 | # The cutoff on the size of index values that determines which filesort algorithm to use. 68 | max_length_for_sort_data = {'type':'infer','min_value': 4,'max_value': 8388608,'length':1} 69 | # This variable limits the total number of prepared statements in the server. Setting the value to 0 disables prepared statements. 70 | max_prepared_stmt_count = {'type':'infer','min_value': 0,'max_value': 1048576,'length':1} 71 | # The number of times that any given stored procedure may be called recursively. The default value for this option is 0, which completely disables recursion in stored procedures. The maximum value is 255. 72 | max_sp_recursion_depth = {'type':'infer','min_value': 0,'max_value': 255,'length':1} 73 | # The maximum number of simultaneous connections permitted to any given MySQL user account. A value of 0 (the default) means “no limit.” 74 | max_user_connections = {'type':'infer','min_value': 0,'max_value': 4294967295,'length':1} 75 | # If this value is greater than 1, MyISAM table indexes are created in parallel. 76 | myisam_repair_threads = {'type':'infer','min_value': 1,'max_value': 4294967295,'length':1} 77 | #for test use ,innodb_buffer_pool_instances&innodb_buffer_pool_chunk_size is read only 78 | read_buffer_size = {'type':'infer','min_value': 1,'max_value': 8589934592,'length':32} 79 | # The size of buffer pool (bytes). A larger buffer pool requires less disk I/O to access the same table data more than once. 80 | innodb_buffer_pool_size = {'type':'infer','min_value': 1048576,'max_value': 8589934592,'length':1048576} -------------------------------------------------------------------------------- /configs.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | 3 | import configparser 4 | import json 5 | 6 | 7 | class DictParser(configparser.ConfigParser): 8 | def read_dict(self): 9 | d = dict(self._sections) 10 | for k in d: 11 | d[k] = dict(d[k]) 12 | return d 13 | 14 | 15 | cf = DictParser() 16 | cf.read("config.ini", encoding="utf-8") 17 | config_dict = cf.read_dict() 18 | 19 | 20 | def parse_args(): 21 | return config_dict["database_tune"] 22 | 23 | 24 | def parse_knob_config(): 25 | _knob_config = config_dict["knob_config"] 26 | for key in _knob_config: 27 | _knob_config[key] = json.loads(str(_knob_config[key]).replace("\'", "\"")) 28 | return _knob_config 29 | 30 | 31 | knob_config = parse_knob_config() 32 | # sync with main.py 33 | 34 | predictor_output_dim = int(config_dict["predictor"]["predictor_output_dim"]) 35 | 36 | predictor_epoch = int(config_dict["predictor"]["predictor_epoch"]) 37 | -------------------------------------------------------------------------------- /draw.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import numpy as np 3 | import matplotlib as mpl 4 | import matplotlib.pyplot as plt 5 | from matplotlib import rcParams 6 | import sys 7 | 8 | 9 | def draw_lines(filelist, metric_name): 10 | ''' 11 | multiple lines on the same metric (y) with increasing iterations (x) 12 | :param filelist: 13 | :param metric_name: 14 | :return: 1 (succeed)/0 (fail) 15 | ''' 16 | 17 | ''' Load Data: [QTune] ''' 18 | col_list = ["iteration", metric_name] 19 | df = pd.read_csv("training-results/" + filelist[0], usecols=col_list, sep="\t") 20 | 21 | x_qtune = list(df[col_list[0]]) 22 | x_qtune = [int(x) for x in x_qtune] 23 | y_qtune = list(df[col_list[1]]) 24 | y_qtune = [float(y) for y in y_qtune] 25 | 26 | ''' Load Data: [Random] ''' 27 | col_list = ["iteration", metric_name] 28 | df = pd.read_csv("training-results/" + filelist[1], usecols=col_list, sep="\t") 29 | x_random = list(df[col_list[0]]) 30 | x_random = [int(x) for x in x_random] 31 | y_random = list(df[col_list[1]]) 32 | y_random = [float(y) for y in y_random] 33 | 34 | ''' figure drawing ''' 35 | mpl.rcdefaults() 36 | rcParams.update({ 37 | 'xtick.labelsize': 12, 38 | 'ytick.labelsize': 12, 39 | 'axes.labelsize': 15, 40 | # 'figure.autolayout': True, 41 | 'figure.subplot.hspace': 0.45, 42 | 'figure.subplot.wspace': 0.22, 43 | # 'mathtext.fontset': 'cm', 44 | }) 45 | 46 | fig = plt.figure() 47 | 48 | qid = 1 49 | ax = fig.add_subplot(1, 1, qid) 50 | 51 | rf = np.array(y_qtune) 52 | dt = np.array(y_random) 53 | 54 | x = np.arange(1, max(len(x_random), len(x_qtune)) + 5) 55 | # y = np.arange(0.0, 1.0) 56 | 57 | l1, = plt.plot(x_qtune, rf[:len(x_qtune)], marker='D', ms=3, linewidth=1) 58 | l2, = plt.plot(x_random, dt[:len(x_random)], marker='X', ms=3, linewidth=1) 59 | 60 | ax.text(0.5, -0.36, 61 | f"({chr(ord('a') + qid - 1)}) $D_{{ {qid} }}$", 62 | horizontalalignment='center', transform=ax.transAxes, fontsize=15, family='serif', 63 | ) 64 | ax.set_xticks(np.arange(0, len(x), len(x) / 10)) 65 | if metric_name == 'latency': 66 | y_range = max(max(y_qtune), max(y_random)) + 5 67 | elif metric_name == 'throughput': 68 | y_range = max(max(y_qtune), max(y_random)) + 100 69 | 70 | ax.set_yticks(np.arange(0, y_range, y_range / 10)) 71 | ax.set_ylim(0, y_range) 72 | ax.set_xlim(0, len(x)) 73 | ax.set_xlabel('#-Iterations') 74 | ax.set_ylabel('Performance') 75 | 76 | fig.legend([l1, l2], ['QTune', 'Random'], 77 | loc='upper center', ncol=4, 78 | handlelength=3, 79 | columnspacing=6., 80 | bbox_to_anchor=(0., 0.98, 1., .0), 81 | bbox_transform=plt.gcf().transFigure, 82 | fontsize=10, 83 | ) 84 | 85 | plt.savefig('training-results/training.png') 86 | 87 | return 1 88 | 89 | 90 | if __name__ == '__main__': 91 | argv = sys.argv 92 | linelist = argv[1].split(',') 93 | metric_name = argv[2] 94 | mark = draw_lines(linelist, metric_name) 95 | if mark: 96 | print('Successfully update figure!') 97 | else: 98 | print('Fail to update figure!') 99 | -------------------------------------------------------------------------------- /environment.py: -------------------------------------------------------------------------------- 1 | import datetime 2 | import subprocess 3 | from collections import deque 4 | import numpy as np 5 | 6 | import pymysql 7 | import pymysql.cursors as pycursor 8 | 9 | import gym 10 | from gym import spaces 11 | from gym.utils import seeding 12 | 13 | from sql2resource import SqlParser 14 | 15 | from configs import knob_config 16 | import time 17 | from run_job import run_job 18 | 19 | 20 | # fetch all the knobs from the prepared configuration info 21 | 22 | 23 | class Database: 24 | def __init__(self, argus): 25 | self.argus = argus 26 | # self.internal_metric_num = 13 # 13(state) + cumulative() 27 | self.external_metric_num = 2 # [throughput, latency] # num_event / t 28 | self.max_connections_num = None 29 | self.knob_names = [knob for knob in knob_config] 30 | print("knob_names:", self.knob_names) 31 | self.knob_num = len(knob_config) 32 | self.internal_metric_num = 65 # default system metrics enabled in metric_innodb 33 | self.max_connections() 34 | try: 35 | conn = self._get_conn() 36 | cursor = conn.cursor() 37 | sql = "SELECT count FROM INFORMATION_SCHEMA.INNODB_METRICS where status='enabled'" 38 | cursor.execute(sql) 39 | result = cursor.fetchall() 40 | self.internal_metric_num = len(result) 41 | cursor.close() 42 | conn.close() 43 | except Exception as err: 44 | print("execute sql error:", err) 45 | 46 | def _get_conn(self): 47 | conn = pymysql.connect(host=self.argus['host'], 48 | port=int(self.argus['port']), 49 | user=self.argus['user'], 50 | password=self.argus['password'], 51 | db='INFORMATION_SCHEMA', 52 | connect_timeout=36000, 53 | cursorclass=pycursor.DictCursor) 54 | return conn 55 | 56 | def fetch_internal_metrics(self): 57 | ######### observation_space 58 | # State_status 59 | # [lock_row_lock_time_max, lock_row_lock_time_avg, buffer_pool_size, 60 | # buffer_pool_pages_total, buffer_pool_pages_misc, buffer_pool_pages_data, buffer_pool_bytes_data, 61 | # buffer_pool_pages_dirty, buffer_pool_bytes_dirty, buffer_pool_pages_free, trx_rseg_history_len, 62 | # file_num_open_files, innodb_page_size] 63 | # Cumulative_status 64 | # [lock_row_lock_current_waits, ] 65 | ''' 66 | sql = "select count from INNODB_METRICS where name='lock_row_lock_time_max' or name='lock_row_lock_time_avg'\ 67 | or name='buffer_pool_size' or name='buffer_pool_pages_total' or name='buffer_pool_pages_misc' or name='buffer_pool_pages_data'\ 68 | or name='buffer_pool_bytes_data' or name='buffer_pool_pages_dirty' or name='buffer_pool_bytes_dirty' or name='buffer_pool_pages_free'\ 69 | or name='trx_rseg_history_len' or name='file_num_open_files' or name='innodb_page_size'" 70 | ''' 71 | state_list = np.array([]) 72 | try: 73 | conn = self._get_conn() 74 | cursor = conn.cursor() 75 | sql = "SELECT count FROM INFORMATION_SCHEMA.INNODB_METRICS where status='enabled'" 76 | cursor.execute(sql) 77 | result = cursor.fetchall() 78 | for s in result: 79 | state_list = np.append(state_list, [s['count']]) 80 | cursor.close() 81 | conn.close() 82 | except Exception as error: 83 | print(error) 84 | 85 | return state_list 86 | 87 | def fetch_knob(self): 88 | state_list = np.append([], []) 89 | try: 90 | conn = self._get_conn() 91 | cursor = conn.cursor() 92 | sql = "select" 93 | for i, knob in enumerate(self.knob_names): 94 | sql = sql + ' @@' + knob 95 | if i < self.knob_num - 1: 96 | sql = sql + ', ' 97 | # print("fetch_knob:", sql) 98 | cursor.execute(sql) 99 | result = cursor.fetchall() 100 | for i in range(self.knob_num): 101 | state_list = np.append(state_list, result[0]["@@%s" % self.knob_names[i]]) 102 | cursor.close() 103 | conn.close() 104 | except Exception as error: 105 | print("fetch_knob Error:", error) 106 | return state_list 107 | 108 | def max_connections(self): 109 | # if not self.max_connections_num: 110 | if 1: 111 | try: 112 | conn = self._get_conn() 113 | cursor = conn.cursor() 114 | sql = "show global variables like 'max_connections';" 115 | cursor.execute(sql) 116 | self.max_connections_num = int(cursor.fetchone()["Value"]) 117 | cursor.close() 118 | conn.close() 119 | except Exception as error: 120 | print(error) 121 | return self.max_connections_num 122 | 123 | def change_knob_nonrestart(self, actions): 124 | try: 125 | conn = self._get_conn() 126 | for i in range(self.knob_num): 127 | cursor = conn.cursor() 128 | if self.knob_names[i] == 'max_connections': 129 | self.max_connections_num = actions[i] 130 | sql = 'set global %s=%d' % (self.knob_names[i], actions[i]) 131 | cursor.execute(sql) 132 | # print(f"修改参数-{self.knob_names[i]}:{actions[i]}") 133 | conn.commit() 134 | conn.close() 135 | return 1 136 | except Exception as error: 137 | conn.close() 138 | print("change_knob_nonrestart error：", error) 139 | return 0 140 | 141 | 142 | # Define the environment 143 | class Environment(gym.Env): 144 | 145 | def __init__(self, db, argus): 146 | 147 | self.db = db 148 | 149 | self.parser = SqlParser(argus) 150 | 151 | self.state_num = db.internal_metric_num 152 | self.action_num = db.knob_num 153 | self.timestamp = int(time.time()) 154 | 155 | # pfs = open('training-results/res_predict-' + str(self.timestamp), 'a') 156 | # pfs.write("%s\t%s\t%s\n" % ('iteration', 'throughput', 'latency')) 157 | # pfs.close() 158 | # 159 | # rfs = open('training-results/res_random-' + str(self.timestamp), 'a') 160 | # rfs.write("%s\t%s\t%s\n" % ('iteration', 'throughput', 'latency')) 161 | # rfs.close() 162 | 163 | ''' observation dim = system metric dim + query vector dim ''' 164 | self.score = 0 # accumulate rewards 165 | 166 | self.o_dim = db.internal_metric_num + len(self.db.fetch_internal_metrics()) 167 | self.o_low = np.array([-1e+10] * self.o_dim) 168 | self.o_high = np.array([1e+10] * self.o_dim) 169 | 170 | self.observation_space = spaces.Box(low=self.o_low, high=self.o_high, dtype=np.float32) 171 | # part 1: current system metric 172 | self.state = db.fetch_internal_metrics() 173 | # print("Concatenated state:") 174 | # part 2: predicted system metric after executing the workload 175 | self.workload = argus["workload"] 176 | 177 | # TODO: 打开训练predict的方法后，此方法注释 178 | ################################################################################ 179 | state0 = self.db.fetch_internal_metrics() 180 | self.preheat() 181 | state1 = self.db.fetch_internal_metrics() 182 | try: 183 | if self.parser.predict_sql_resource_value is None: 184 | self.parser.predict_sql_resource_value = state1 - state0 185 | except Exception as error: 186 | print("get predict_sql_resource_value error:", error) 187 | ################################################################################ 188 | 189 | self.state = np.append(self.parser.predict_sql_resource(self.workload), self.state) 190 | 191 | ''' action space ''' 192 | # Offline 193 | # table_open_cache(1), max_connections(2), innodb_buffer_pool_instances(4), 194 | # innodb_log_files_in_group(5), innodb_log_file_size(6), innodb_purge_threads(7), innodb_read_io_threads(8) 195 | # innodb_write_io_threads(9), 196 | # Online 197 | # innodb_buffer_pool_size(3), max_binlog_cache_size(10), binlog_cache_size(11) 198 | # 1 2 3 11 199 | # exclude 200 | # innodb_file_per_table, skip_name_resolve, binlog_checksum, 201 | # binlog_format(dynamic, [ROW, STATEMENT, MIXED]), 202 | 203 | calculate_knobs = [] 204 | infer_knobs = [] 205 | for k in knob_config.items(): 206 | if k[1]['type'] == 'infer': 207 | infer_knobs.append(k) 208 | else: 209 | calculate_knobs.append(k) 210 | self.knob_num = len(knob_config) 211 | # self.a_low = np.array([knob[1]['min_value']/knob[1]['length'] for knob in list(knob_config.items())[:db.knob_num]]) 212 | self.a_low = np.array([knob[1]['min_value'] / knob[1]['length'] for knob in infer_knobs]) 213 | # self.a_high = np.array([knob[1]['max_value']/knob[1]['length'] for knob in list(knob_config.items())[:db.knob_num]]) 214 | self.a_high = np.array([knob[1]['max_value'] / knob[1]['length'] for knob in infer_knobs]) 215 | # self.length = np.array([knob[1]['length'] for knob in list(knob_config.items())[:db.knob_num]]) 216 | self.length = np.array([knob[1]['length'] * 1.0 for knob in infer_knobs]) 217 | self.action_space = spaces.Box(low=self.a_low, high=self.a_high, dtype=np.float32) 218 | self.default_action = self.a_low 219 | self.mem = deque(maxlen=int(argus['maxlen_mem'])) # [throughput, latency] 220 | self.predicted_mem = deque(maxlen=int(argus['maxlen_predict_mem'])) 221 | self.knob2pos = {knob: i for i, knob in enumerate(knob_config)} 222 | self.seed() 223 | self.start_time = datetime.datetime.now() 224 | 225 | def seed(self, seed=None): 226 | self.np_random, seed = seeding.np_random(seed) 227 | return [seed] 228 | 229 | def execute_command(self): 230 | self.db.max_connections() 231 | # print(self.db.max_connections_num) 232 | if self.parser.argus['thread_num_auto'] == '0': 233 | thread_num = int(self.parser.argus['thread_num']) 234 | else: 235 | thread_num = int(self.db.max_connections_num) - 1 236 | run_job(thread_num, self.workload, self.parser.resfile) 237 | 238 | def preheat(self): 239 | self.execute_command() 240 | 241 | def fetch_action(self): 242 | while True: 243 | state_list = self.db.fetch_knob() 244 | if list(state_list): 245 | break 246 | time.sleep(5) 247 | return state_list 248 | 249 | # new_state, reward, done, 250 | def step(self, u, isPredicted, iteration, action_tmp=None): 251 | flag = self.db.change_knob_nonrestart(u) 252 | 253 | # if failing to tune knobs, give a high panlty 254 | if not flag: 255 | return self.state, -10e+4, self.score, 1 256 | 257 | self.execute_command() 258 | throughput, latency = self._get_throughput_latency() 259 | # ifs = open('fl1', 'r') 260 | # print(str(len(self.mem)+1)+"\t"+str(throughput)+"\t"+str(latency)) 261 | cur_time = datetime.datetime.now() 262 | interval = (cur_time - self.start_time).seconds 263 | self.mem.append([throughput, latency]) 264 | # 2 refetch state 265 | self._get_obs() 266 | # 3 cul reward(T, L) 267 | reward = self._calculate_reward(throughput, latency) 268 | ''' 269 | reward = 0 270 | for i in range(u.shape[0]): 271 | tmp = u[i] / self.a_high[i] 272 | reward+=tmp 273 | # print("Performance: %d\t%f\t%f\t%f\t%ds" % (len(self.mem) + 1, throughput, latency, reward, interval)) 274 | if isPredicted: 275 | self.predicted_mem.append([len(self.predicted_mem), throughput, latency, reward]) 276 | if len(self.predicted_mem)%10 == 0: 277 | print("Predict List") 278 | print(self.predicted_mem) 279 | ''' 280 | 281 | action = self.fetch_action() 282 | 283 | if isPredicted: 284 | self.predicted_mem.append([len(self.predicted_mem), throughput, latency, reward]) 285 | 286 | # print("Predict %d\t%f\t%f\t%f\t%ds" % (len(self.mem) + 1, throughput, latency, reward, interval)) 287 | 288 | pfs = open('training-results/res_predict-' + str(self.timestamp), 'a') 289 | pfs.write("%d\t%f\t%f\n" % (iteration, throughput, latency)) 290 | pfs.close() 291 | 292 | fetch_knob = open('training-results/fetch_knob_predict-' + str(self.timestamp), 'a') 293 | fetch_knob.write(f"{str(iteration)}\t{str(list(action))}\n") 294 | fetch_knob.close() 295 | 296 | action_write = open('training-results/action_test_predict-' + str(self.timestamp), 'a') 297 | action_write.write(f"{str(iteration)}\t{str(list(u))}\n") 298 | action_write.write(f"{str(iteration)}\t{str(list(action_tmp))}\n") 299 | action_write.close() 300 | 301 | self.score = self.score + reward 302 | 303 | else: 304 | # print("Random %d\t%f\t%f\t%f\t%ds" % (len(self.mem) + 1, throughput, latency, reward, interval)) 305 | 306 | rfs = open('training-results/res_random-' + str(self.timestamp), 'a') 307 | rfs.write("%d\t%f\t%f\n" % (iteration, throughput, latency)) 308 | rfs.close() 309 | 310 | action_write = open('training-results/action_random-' + str(self.timestamp), 'a') 311 | action_write.write(f"{str(iteration)}\t{str(list(u))}\n") 312 | action_write.close() 313 | 314 | fetch_knob = open('training-results/fetch_knob_random-' + str(self.timestamp), 'a') 315 | fetch_knob.write(f"{str(iteration)}\t{str(list(action))}\n") 316 | fetch_knob.close() 317 | 318 | return self.state, reward, self.score, throughput 319 | 320 | def _get_throughput_latency(self): 321 | with open(self.parser.resfile, 'r') as f: 322 | try: 323 | for line in f.readlines(): 324 | a = line.split() 325 | if len(a) > 1 and 'avg_qps(queries/s):' == a[0]: 326 | throughput = float(a[1]) 327 | if len(a) > 1 and 'avg_lat(s):' == a[0]: 328 | latency = float(a[1]) 329 | finally: 330 | f.close() 331 | # print("throughput:{} \n latency:{}".format(throughput, latency)) 332 | return throughput, latency 333 | 334 | def _calculate_reward(self, throughput, latency): 335 | if len(self.mem) != 0: 336 | dt0 = (throughput - self.mem[0][0]) / self.mem[0][0] 337 | dt1 = (throughput - self.mem[len(self.mem) - 1][0]) / self.mem[len(self.mem) - 1][0] 338 | if dt0 >= 0: 339 | rt = ((1 + dt0) ** 2 - 1) * abs(1 + dt1) 340 | else: 341 | rt = -((1 - dt0) ** 2 - 1) * abs(1 - dt1) 342 | 343 | dl0 = -(latency - self.mem[0][1]) / self.mem[0][1] 344 | 345 | dl1 = -(latency - self.mem[len(self.mem) - 1][1]) / self.mem[len(self.mem) - 1][1] 346 | 347 | if dl0 >= 0: 348 | rl = ((1 + dl0) ** 2 - 1) * abs(1 + dl1) 349 | else: 350 | rl = -((1 - dl0) ** 2 - 1) * abs(1 - dl1) 351 | 352 | else: # initial action 353 | rt = 0 354 | rl = 0 355 | reward = 1 * rl + 9 * rt 356 | return reward 357 | 358 | def _get_obs(self): 359 | self.state = self.db.fetch_internal_metrics() 360 | self.state = np.append(self.parser.predict_sql_resource(self.workload), self.state) 361 | return self.state 362 | -------------------------------------------------------------------------------- /generate_sysbench_insert_and_delete_query.py: -------------------------------------------------------------------------------- 1 | table_count = 20 2 | rows_count = 10000 3 | 4 | 5 | def write_insert_to_file(file_path): 6 | with open(file_path, 'a+') as f: 7 | for i in range(0, int(rows_count / table_count)): 8 | 9 | for j in range(1, table_count + 1): 10 | insert = "insert into sbtest{} (k, c, pad) values ((select floor(2+rand()*100000)), '{}', '94657455071-01886877449-66853068383-97480802197-06448926027');".format( 11 | j, i + 1) 12 | f.write(insert) 13 | f.write('\n') 14 | 15 | f.close() 16 | print("写入成功") 17 | 18 | 19 | def write_delete_to_file(file_path): 20 | with open(file_path, 'a+') as f: 21 | for i in range(0, int(rows_count / table_count)): 22 | for j in range(1, table_count + 1): 23 | delete = "delete from sbtest{} order by id desc limit 1;".format(j) 24 | f.write(delete) 25 | f.write('\n') 26 | f.close() 27 | print("写入成功") 28 | 29 | 30 | # Press the green button in the gutter to run the script. 31 | if __name__ == '__main__': 32 | write_insert_to_file("/Users/Four/Desktop/insert.txt") 33 | write_delete_to_file("/Users/Four/Desktop/delete.txt") 34 | -------------------------------------------------------------------------------- /get_res_data_from_file.py: -------------------------------------------------------------------------------- 1 | def data_from_file(file_path): 2 | indexs = [] 3 | datas = [] 4 | delaies = [] 5 | try: 6 | with open(file_path, 'r') as f: 7 | result = f.read().splitlines() 8 | for i in result: 9 | a = i.split() 10 | indexs.append(a[0]) 11 | datas.append(a[1]) 12 | delaies.append(a[2]) 13 | f.close() 14 | except Exception as error: 15 | print(str(error)) 16 | 17 | print(indexs) 18 | print(datas) 19 | print(delaies) 20 | 21 | return indexs, datas, delaies 22 | 23 | 24 | def knob_data_from_file(file_path): 25 | indexs = [] 26 | datas = [] 27 | try: 28 | with open(file_path, 'r') as f: 29 | result = f.read().splitlines() 30 | for i in result: 31 | a = i.split("\t") 32 | print(a) 33 | indexs.append(a[0]) 34 | datas.append(a[1]) 35 | f.close() 36 | except Exception as error: 37 | print(str(error)) 38 | 39 | print(indexs) 40 | print(datas) 41 | 42 | return indexs, datas 43 | 44 | 45 | if __name__ == '__main__': 46 | data_from_file("/Users/Four/Desktop/qtune_results/res_predict-1625452896") 47 | print("\n\n") 48 | data_from_file("/Users/Four/Desktop/qtune_results/res_random-1625452896") 49 | print("\n\n") 50 | -------------------------------------------------------------------------------- /get_sysbench_workload_from_slow_file.py: -------------------------------------------------------------------------------- 1 | def get_workload(file_path): 2 | workload = [] 3 | try: 4 | with open(file_path, 'r') as f: 5 | result = f.read().splitlines() 6 | for i in result: 7 | if ("#" not in i) and ("SET timestamp" not in i) and ("COMMIT" not in i) and ("BEGIN" not in i) and i: 8 | workload.append(i) 9 | f.close() 10 | print(f"共有{len(workload)}条数据") 11 | with open("/Users/Four/Desktop/workload123.txt", "w+") as sql_file: 12 | for query in workload: 13 | if query: 14 | sql_file.write(query + "\n") 15 | sql_file.close() 16 | except Exception as error: 17 | print(str(error)) 18 | 19 | 20 | if __name__ == '__main__': 21 | 22 | # get_workload("/Users/Four/Desktop/slow.log") 23 | 24 | for i in range(0, 1): 25 | 26 | for j in range(1, 21): 27 | insert = "insert into sbtest{} (k, c, pad) values ((select floor(2+rand()*100000)), '{}', '94657455071-01886877449-66853068383-97480802197-06448926027');".format( 28 | j, i + 1) 29 | print(insert) 30 | -------------------------------------------------------------------------------- /get_workload_from_file.py: -------------------------------------------------------------------------------- 1 | def get_workload_from_file(file_path): 2 | workload = [] 3 | try: 4 | with open(file_path, 'r') as f: 5 | result = f.read().splitlines() 6 | for i in result: 7 | if i: 8 | workload.append(i) 9 | f.close() 10 | except Exception as error: 11 | print(str(error)) 12 | return workload 13 | 14 | -------------------------------------------------------------------------------- /knob-effect-test.py: -------------------------------------------------------------------------------- 1 | import sys 2 | import time 3 | import numpy as np 4 | import tensorflow as tf 5 | import keras.backend as K 6 | from environment import Database, Environment 7 | from model import ActorCritic 8 | from configs import parse_args 9 | from get_workload_from_file import get_workload_from_file 10 | 11 | if __name__ == "__main__": 12 | 13 | argus = parse_args() 14 | 15 | # prepare_training_workloads 16 | training_workloads = [] 17 | workload = get_workload_from_file(argus["workload_file_path"]) 18 | argus["workload"] = workload 19 | sess = tf.Session() 20 | K.set_session(sess) 21 | db = Database(argus) # connector knobs metric 22 | env = Environment(db, argus) 23 | 24 | # TODO: 训练predict 25 | # sample_times = 2 26 | # for i in range(sample_times): 27 | # training_workloads.append(np.random.choice(workload, np.random.randint(len(workload)), replace=False, p=None)) 28 | # X = [] 29 | # Y = [] 30 | # for w in training_workloads: 31 | # vec = env.parser.get_workload_encoding(w) 32 | # X.append(vec.flatten()) 33 | # state0 = env.db.fetch_internal_metrics() 34 | # env.preheat() 35 | # state1 = env.db.fetch_internal_metrics() 36 | # Y.append(state1 - state0) 37 | # X = np.array(X) 38 | # Y = np.array(Y) 39 | # env.parser.estimator.fit(X, Y, batch_size=50, epochs=predictor_epoch) 40 | 41 | # TODO save&load model e.g. env.parser.estimator.save_weights(path) 42 | # env.parser.estimator.save_weights(filepath=path) 43 | # env.parser.estimator.load_weights(filepath=path) 44 | 45 | actor_critic = ActorCritic(env, sess, learning_rate=float(argus['learning_rate']), 46 | train_min_size=int(argus['train_min_size']), 47 | size_mem=int(argus['maxlen_mem']), size_predict_mem=int(argus['maxlen_predict_mem'])) 48 | 49 | num_trials = int(argus['num_trial']) # ? 50 | # trial_len = 500 # ? 51 | # ntp 52 | 53 | interval = 1 54 | time.sleep(interval) 55 | for value in range(3,12,1): 56 | sql = 'set global %s=%d' % ('max_connections', value) 57 | conn = db._get_conn() 58 | cursor = conn.cursor() 59 | cursor.execute(sql) 60 | print(f"修改参数-max_connections:{value}") 61 | conn.commit() 62 | conn.close() 63 | 64 | env.execute_command() 65 | time.sleep(interval) 66 | 67 | 68 | exit() 69 | 70 | # First iteration 71 | cur_state = env._get_obs() # np.array (inner_metric + sql) 72 | cur_state = cur_state.reshape((1, env.state.shape[0])) 73 | # action = env.action_space.sample() 74 | action = env.fetch_action() # np.array 75 | action_2 = action.reshape((1, env.knob_num)) # for memory 76 | action_2 = action_2[:, :env.action_space.shape[0]] 77 | new_state, reward, socre, cur_throughput = env.step(action, 0, 78 | 1) # apply the action -> to steady state -> return the reward 79 | new_state = new_state.reshape((1, env.state.shape[0])) 80 | reward_np = np.array([reward]) 81 | print(reward_np) 82 | actor_critic.remember(cur_state, action_2, reward_np, new_state, False) 83 | actor_critic.train(1) # len<32, useless 84 | 85 | cur_state = new_state 86 | predicted_rewardList = [] 87 | for epoch in range(num_trials): 88 | # env.render() 89 | cur_state = cur_state.reshape((1, env.state.shape[0])) 90 | action, isPredicted, action_tmp = actor_critic.act(cur_state) 91 | # action.tolist() # to execute 92 | new_state, reward, score, throughput = env.step(action, isPredicted, epoch + 1, action_tmp) 93 | new_state = new_state.reshape((1, env.state.shape[0])) 94 | 95 | action = env.fetch_action() 96 | action_2 = action.reshape((1, env.knob_num)) # for memory 97 | action_2 = action_2[:, :env.action_space.shape[0]] 98 | 99 | if isPredicted == 1: 100 | predicted_rewardList.append([epoch, reward]) 101 | 102 | reward_np = np.array([reward]) 103 | 104 | actor_critic.remember(cur_state, action_2, reward_np, new_state, False) 105 | actor_critic.train(epoch) 106 | 107 | print('============train running==========') 108 | 109 | if epoch % 5 == 0: 110 | print('============save_weights==========') 111 | actor_critic.actor_model.save_weights('saved_model_weights/actor_weights.h5') 112 | actor_critic.critic_model.save_weights('saved_model_weights/critic_weights.h5') 113 | 114 | if (throughput - cur_throughput) / cur_throughput > float(argus['stopping_throughput_improvement_percentage']): 115 | print("training end!!") 116 | env.parser.close_mysql_conn() 117 | break 118 | 119 | cur_state = new_state 120 | 121 | 122 | 123 | 124 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | import sys 2 | import time 3 | import numpy as np 4 | import tensorflow as tf 5 | import keras.backend as K 6 | from environment import Database, Environment 7 | from model import ActorCritic 8 | from configs import parse_args 9 | from get_workload_from_file import get_workload_from_file 10 | 11 | if __name__ == "__main__": 12 | 13 | argus = parse_args() 14 | 15 | # prepare_training_workloads 16 | training_workloads = [] 17 | workload = get_workload_from_file(argus["workload_file_path"]) 18 | argus["workload"] = workload 19 | sess = tf.Session() 20 | K.set_session(sess) 21 | db = Database(argus) # connector knobs metric 22 | env = Environment(db, argus) 23 | 24 | # TODO: 训练predict 25 | # sample_times = 2 26 | # for i in range(sample_times): 27 | # training_workloads.append(np.random.choice(workload, np.random.randint(len(workload)), replace=False, p=None)) 28 | # X = [] 29 | # Y = [] 30 | # for w in training_workloads: 31 | # vec = env.parser.get_workload_encoding(w) 32 | # X.append(vec.flatten()) 33 | # state0 = env.db.fetch_internal_metrics() 34 | # env.preheat() 35 | # state1 = env.db.fetch_internal_metrics() 36 | # Y.append(state1 - state0) 37 | # X = np.array(X) 38 | # Y = np.array(Y) 39 | # env.parser.estimator.fit(X, Y, batch_size=50, epochs=predictor_epoch) 40 | 41 | # TODO save&load model e.g. env.parser.estimator.save_weights(path) 42 | # env.parser.estimator.save_weights(filepath=path) 43 | # env.parser.estimator.load_weights(filepath=path) 44 | 45 | actor_critic = ActorCritic(env, sess, learning_rate=float(argus['learning_rate']), 46 | train_min_size=int(argus['train_min_size']), 47 | size_mem=int(argus['maxlen_mem']), size_predict_mem=int(argus['maxlen_predict_mem'])) 48 | 49 | num_trials = int(argus['num_trial']) # ? 50 | # trial_len = 500 # ? 51 | # ntp 52 | 53 | 54 | # First iteration 55 | cur_state = env._get_obs() # np.array (inner_metric + sql) 56 | cur_state = cur_state.reshape((1, env.state.shape[0])) 57 | # action = env.action_space.sample() 58 | action = env.fetch_action() # np.array 59 | action_2 = action.reshape((1, env.knob_num)) # for memory 60 | action_2 = action_2[:, :env.action_space.shape[0]] 61 | new_state, reward, socre, cur_throughput = env.step(action, 0, 62 | 1) # apply the action -> to steady state -> return the reward 63 | new_state = new_state.reshape((1, env.state.shape[0])) 64 | reward_np = np.array([reward]) 65 | print(reward_np) 66 | actor_critic.remember(cur_state, action_2, reward_np, new_state, False) 67 | actor_critic.train(1) # len<[train_min_size], useless 68 | 69 | cur_state = new_state 70 | predicted_rewardList = [] 71 | for epoch in range(num_trials): 72 | # env.render() 73 | cur_state = cur_state.reshape((1, env.state.shape[0])) 74 | action, isPredicted, action_tmp = actor_critic.act(cur_state) 75 | # action.tolist() # to execute 76 | new_state, reward, score, throughput = env.step(action, isPredicted, epoch + 1, action_tmp) 77 | new_state = new_state.reshape((1, env.state.shape[0])) 78 | 79 | action = env.fetch_action() 80 | action_2 = action.reshape((1, env.knob_num)) # for memory 81 | action_2 = action_2[:, :env.action_space.shape[0]] 82 | 83 | if isPredicted == 1: 84 | predicted_rewardList.append([epoch, reward]) 85 | print("[predicted]", action_2, reward, throughput) 86 | else: 87 | print("[random]", action_2, reward, throughput) 88 | 89 | reward_np = np.array([reward]) 90 | 91 | actor_critic.remember(cur_state, action_2, reward_np, new_state, False) 92 | actor_critic.train(epoch) 93 | 94 | # print('============train running==========') 95 | 96 | if epoch % 5 == 0: 97 | # print('============save_weights==========') 98 | actor_critic.actor_model.save_weights('saved_model_weights/actor_weights.h5') 99 | actor_critic.critic_model.save_weights('saved_model_weights/critic_weights.h5') 100 | ''' 101 | if (throughput - cur_throughput) / cur_throughput > float(argus['stopping_throughput_improvement_percentage']): 102 | print("training end!!") 103 | env.parser.close_mysql_conn() 104 | break 105 | ''' 106 | 107 | cur_state = new_state 108 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import sys 2 | import datetime 3 | from os import path 4 | import subprocess 5 | import time 6 | from collections import deque 7 | import numpy as np 8 | import random 9 | import tensorflow as tf 10 | import pandas 11 | import heapq 12 | import pymysql 13 | import pymysql.cursors as pycursor 14 | 15 | import gym 16 | from gym import spaces 17 | from gym.utils import seeding 18 | 19 | import os 20 | from keras.models import Sequential, Model 21 | from keras.layers import Dense, Dropout, Input 22 | from keras.layers.merge import Add, Multiply 23 | from keras.optimizers import Adam 24 | import keras.backend as K 25 | from keras.layers.normalization import BatchNormalization 26 | from keras.wrappers.scikit_learn import KerasRegressor 27 | from sklearn.preprocessing import StandardScaler 28 | from configs import knob_config 29 | from keras.initializers import random_uniform,ones,constant 30 | 31 | 32 | # determines how to assign values to each state, i.e. takes the state 33 | # and action (two-input model) and determines the corresponding value 34 | # Tunable parameters 35 | # learning_rate = 0.001 36 | # epsilon = 1.0 37 | # epsilon_decay = .995 38 | # gamma = .95 39 | # tau = .125 40 | # 4*relu 41 | class ActorCritic: 42 | def __init__(self, env, sess, learning_rate=0.001, train_min_size=32, size_mem=2000, size_predict_mem=2000): 43 | self.env = env 44 | 45 | self.sess = sess 46 | self.learning_rate = learning_rate # 0.001 47 | self.train_min_size = train_min_size 48 | self.epsilon = .9 49 | self.epsilon_decay = .999 50 | self.gamma = .095 51 | self.tau = .125 52 | self.timestamp = int(time.time()) 53 | # ===================================================================== # 54 | # Actor Model # 55 | # Chain rule: find the gradient of chaging the actor network params in # 56 | # getting closest to the final value network predictions, i.e. de/dA # 57 | # Calculate de/dA as = de/dC * dC/dA, where e is error, C critic, A act # 58 | # ===================================================================== # 59 | self.memory = deque(maxlen=size_mem) 60 | self.mem_predicted = deque(maxlen=size_predict_mem) 61 | self.actor_state_input, self.actor_model = self.create_actor_model() 62 | _, self.target_actor_model = self.create_actor_model() 63 | 64 | self.actor_critic_grad = tf.placeholder(tf.float32, 65 | [None, self.env.action_space.shape[ 66 | 0]]) # where we will feed de/dC (from critic) 67 | 68 | # load pre-trained models 69 | # if os.path.exists('saved_model_weights/actor_weights.h5'): 70 | # self.actor_model.load_weights('saved_model_weights/actor_weights.h5') 71 | # self.target_actor_model.load_weights('saved_model_weights/actor_weights.h5') 72 | 73 | actor_model_weights = self.actor_model.trainable_weights 74 | self.actor_grads = tf.gradients(self.actor_model.output, 75 | actor_model_weights, -self.actor_critic_grad) # dC/dA (from actor) 76 | grads = zip(self.actor_grads, actor_model_weights) 77 | self.optimize = tf.train.AdamOptimizer(self.learning_rate).apply_gradients(grads) 78 | 79 | # ===================================================================== # 80 | # Critic Model # 81 | # ===================================================================== # 82 | 83 | self.critic_state_input, self.critic_action_input, \ 84 | self.critic_model = self.create_critic_model() 85 | _, _, self.target_critic_model = self.create_critic_model() 86 | 87 | #if os.path.exists('saved_model_weights/critic_weights.h5'): 88 | # self.critic_model.load_weights('saved_model_weights/critic_weights.h5') 89 | # self.target_critic_model.load_weights('saved_model_weights/critic_weights.h5') 90 | 91 | # print('de:', self.critic_model.output) 92 | # print('dC:', self.critic_action_input) 93 | 94 | self.critic_grads = tf.gradients(self.critic_model.output, 95 | self.critic_action_input) # where we calcaulte de/dC for feeding above 96 | 97 | # Initialize for later gradient calculations 98 | self.sess.run(tf.initialize_all_variables()) 99 | 100 | # ========================================================================= # 101 | # Model Definitions # 102 | # ========================================================================= # 103 | 104 | def create_actor_model(self): 105 | def target_range(x, target_min=self.env.a_low, target_max=self.env.a_high): 106 | x02 = K.tanh(x) + 1 # x in range(0,2) 107 | scale = (target_max - target_min) / 2. 108 | return x02 * scale + target_min 109 | 110 | #def target_range(x, target_min=self.env.a_low, target_max=self.env.a_high): 111 | # scale = (target_max - target_min) 112 | # return scale * K.sigmoid(x) + target_min 113 | 114 | state_input = Input(shape=self.env.observation_space.shape) 115 | 116 | h1 = Dense(128,name = 'h1', activation='relu')(state_input) 117 | n1 =BatchNormalization(axis=1,center=False,scale=False,name='n1')(h1) 118 | h2 = Dense(64, name = 'h2',activation='tanh')(n1) 119 | d1 = Dropout(0.3)(h2) 120 | # add a dense-tanh expend the space!! 121 | #n1 = BatchNormalization(name='n1',center=False,scale=False)(d1) 122 | output = Dense(self.env.action_space.shape[0],activation=target_range)(d1) 123 | 124 | model = Model(input=state_input, output=output) 125 | adam = Adam(lr=0.001) 126 | model.compile(loss="mse", optimizer=adam) 127 | 128 | return state_input, model 129 | 130 | def create_critic_model(self): 131 | # (dense dense)->dense->dense->BN->dense 132 | 133 | state_input = Input(shape=self.env.observation_space.shape) 134 | state_h1 = Dense(128)(state_input) 135 | # state_h2 = Dense(13)(state_h1) 136 | 137 | action_input = Input(shape=self.env.action_space.shape) 138 | action_h1 = Dense(128)(action_input) # 139 | 140 | merged = Add()([state_h1, action_h1]) 141 | merged_h1 = Dense(int(256))(merged) 142 | h2 = Dense(256)(merged_h1) 143 | n1 = BatchNormalization()(h2) 144 | h3 = Dense(64,activation='tanh')(n1) 145 | d1 = Dropout(0.3)(h3) 146 | n1 = BatchNormalization()(d1) 147 | output = Dense(1)(n1) 148 | 149 | model = Model(input=[state_input, action_input], output=output) 150 | 151 | adam = Adam(lr=0.001) 152 | model.compile(loss="mse", optimizer=adam,metrics=['mse']) 153 | return state_input, action_input, model 154 | 155 | # ========================================================================= # 156 | # Model Training # 157 | # ========================================================================= # 158 | 159 | def remember(self, cur_state, action, reward, new_state, done): 160 | self.memory.append([cur_state, action, reward, new_state, done]) 161 | # print("Mem: Q-%f"%reward) 162 | 163 | def _train_actor(self, samples, i): 164 | for sample in samples: 165 | cur_state, action, reward, new_state, _ = sample 166 | cur_state = (cur_state - min(cur_state[0]))/(max(cur_state[0])-min(cur_state[0])) 167 | predicted_action = self.actor_model.predict(cur_state) 168 | h1 = Model(self.actor_model.input,self.actor_model.get_layer('h1').output) 169 | h2 = Model(self.actor_model.input, self.actor_model.get_layer('h2').output) 170 | n1 = Model(self.actor_model.input,self.actor_model.get_layer('n1').output) 171 | print('predicted_action'*5) 172 | print(predicted_action) 173 | print(h1.predict(cur_state)) 174 | print(h2.predict(cur_state)) 175 | res_n1 = n1.predict(cur_state)[0] 176 | print(res_n1) 177 | print(np.mean(res_n1)) 178 | print(np.std(res_n1)) 179 | # print("predicted action", predicted_action) 180 | grads = self.sess.run(self.critic_grads, feed_dict={ 181 | self.critic_state_input: cur_state, 182 | self.critic_action_input: predicted_action 183 | })[0] 184 | # print("first gradient",grads) 185 | self.sess.run(self.optimize, feed_dict={ 186 | self.actor_state_input: cur_state, 187 | self.actor_critic_grad: grads 188 | }) 189 | writer = open('training-results/training-' + str(self.timestamp), 'a') 190 | # writer.write(f"{str(i)}\t{str(list(self.actor_critic_grad))}\n") 191 | writer.write('grads') 192 | writer.write(f"{str(i)}\t{str(list(grads))}\n") 193 | writer.write('cur_state\n') 194 | writer.write(str(cur_state)+'\n') 195 | writer.write('predicted_action\n') 196 | writer.write(str(predicted_action)+'\n') 197 | writer.close() 198 | 199 | def _train_critic(self, samples,i): 200 | for sample in samples: 201 | cur_state, action, t_reward, new_state, done = sample 202 | reward = np.array([]) 203 | reward = np.append(reward, t_reward[0]) 204 | cur_state = (cur_state - min(cur_state[0]))/(max(cur_state[0])-min(cur_state[0])) 205 | # print("<>Q-value:") 206 | # print(reward) 207 | # if not done: 208 | target_action = self.target_actor_model.predict(new_state) 209 | future_reward = self.target_critic_model.predict( 210 | [new_state, target_action])[0][0] 211 | print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~") 212 | print("future_reward:", future_reward) 213 | reward += self.gamma * future_reward 214 | print("reward:", reward) 215 | print("target_action:",target_action) 216 | print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~") 217 | # There comes the convert 218 | # print("Look:") 219 | # print(cur_state.shape) 220 | # print(action.shape) 221 | # print(reward.shape) 222 | # print(reward) 223 | loss = self.critic_model.fit([cur_state, action], reward, verbose=1) # update the Q-value 224 | writer = open('training-results/critic_training-' + str(self.timestamp), 'a') 225 | writer.write('epoch:\t'+str(i)+'\n') 226 | writer.write('critic_loss\t') 227 | writer.write(f"{str(loss.history['loss'])}\n") 228 | writer.write('reward:\t') 229 | writer.write(f"{str(reward)}\n") 230 | writer.close() 231 | def train(self, i): 232 | self.batch_size = self.train_min_size # 32 233 | if len(self.memory) < self.batch_size: 234 | return 235 | mem = list(self.memory) 236 | rewards = [i[2][0] for i in mem] 237 | indexs = heapq.nlargest(self.batch_size, range(len(rewards)), rewards.__getitem__) 238 | samples = [] 239 | for i in indexs: 240 | samples.append(mem[i]) 241 | samples = random.sample(list(self.memory), self.batch_size - 2) 242 | writer = open('training-results/training-' + str(self.timestamp), 'a') 243 | writer.write('samples\n') 244 | writer.write(f"{str(i)}\t{str(np.array(samples)[:,2])}\n") 245 | writer.close() 246 | # print(samples) 247 | self._train_critic(samples,i) 248 | self._train_actor(samples, i) 249 | self.update_target() 250 | 251 | # ========================================================================= # 252 | # Target Model Updating # 253 | # ========================================================================= # 254 | 255 | def _update_actor_target(self): 256 | actor_model_weights = self.actor_model.get_weights() 257 | actor_target_weights = self.target_actor_model.get_weights() 258 | 259 | for i in range(len(actor_target_weights)): 260 | actor_target_weights[i] = actor_model_weights[i] 261 | self.target_actor_model.set_weights(actor_target_weights) 262 | 263 | def _update_critic_target(self): 264 | critic_model_weights = self.critic_model.get_weights() 265 | critic_target_weights = self.target_critic_model.get_weights() 266 | 267 | for i in range(len(critic_target_weights)): 268 | critic_target_weights[i] = critic_model_weights[i] 269 | self.target_critic_model.set_weights(critic_target_weights) 270 | 271 | def update_target(self): 272 | self._update_actor_target() 273 | self._update_critic_target() 274 | 275 | # ========================================================================= # 276 | # Model Predictions # 277 | # ========================================================================= # 278 | 279 | def get_calculate_knobs(self, action): 280 | caculate_knobs = list(knob_config)[len(action):] 281 | for k in caculate_knobs: 282 | 283 | if knob_config[k]['operator'] == 'multiply': 284 | pos_x = self.env.knob2pos[knob_config[k]['x']] 285 | pos_y = self.env.knob2pos[knob_config[k]['y']] 286 | tmp = action[pos_x] * action[pos_y] 287 | action = np.append(action, tmp) 288 | return action 289 | 290 | def act(self, cur_state): 291 | self.epsilon *= self.epsilon_decay 292 | action_tmp = None 293 | if np.random.random(1) < self.epsilon or len(self.memory) < self.batch_size: 294 | print("[Random Tuning]") 295 | action = np.round(self.env.action_space.sample()) 296 | action = action.astype(np.float64) 297 | flag = 0 298 | else: 299 | print("[Model Tuning]") 300 | # action = np.round(self.actor_model.predict(cur_state)[0]) 301 | cur_state = (cur_state - min(cur_state[0]))/(max(cur_state[0])-min(cur_state[0])) 302 | action = self.actor_model.predict(cur_state)[0] 303 | print(action) 304 | # TODO: 临时参数，查看状态使用 305 | action_tmp = action 306 | action = np.round(action) 307 | action = action.astype(np.float64) 308 | flag = 1 309 | 310 | for i in range(action.shape[0]): 311 | if action[i] <= self.env.default_action[i]: 312 | print("[Action %d] Lower than DEFAULT: %f" % (i, action[i])) 313 | action[i] = int(self.env.default_action[i]) * int(self.env.length[i]) 314 | elif action[i] > self.env.a_high[i]: 315 | print("[Action %d] Higher than MAX: %f" % (i, action[i])) 316 | action[i] = int(self.env.a_high[i]) * int(self.env.length[i]) 317 | else: 318 | action[i] = action[i] * self.env.length[i] 319 | 320 | action = self.get_calculate_knobs(action) 321 | 322 | return action, flag, action_tmp 323 | 324 | -------------------------------------------------------------------------------- /pre-processing/knob_rankiing.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from sklearn.linear_model import lasso_path 3 | 4 | ''' 5 | from OtterTune - lasso.py 6 | https://github.com/cmu-db/ottertune/edit/master/server/analysis/lasso.py 7 | ''' 8 | 9 | class LassoPath(): 10 | """Lasso: 11 | 12 | Computes the Lasso path using Sklearn's lasso_path method. 13 | 14 | 15 | See also 16 | -------- 17 | http://scikit-learn.org/stable/modules/generated/sklearn.linear_model.lasso_path.html 18 | 19 | 20 | Attributes 21 | ---------- 22 | feature_labels_ : array, [n_features] 23 | Labels for each of the features in X. 24 | 25 | alphas_ : array, [n_alphas] 26 | The alphas along the path where models are computed. (These are 27 | the decreasing values of the penalty along the path). 28 | 29 | coefs_ : array, [n_outputs, n_features, n_alphas] 30 | Coefficients along the path. 31 | 32 | rankings_ : array, [n_features] 33 | The average ranking of each feature across all target values. 34 | """ 35 | def __init__(self): 36 | self.feature_labels_ = None 37 | self.alphas_ = None 38 | self.coefs_ = None 39 | self.rankings_ = None 40 | 41 | def _reset(self): 42 | """Resets all attributes (erases the model)""" 43 | self.feature_labels_ = None 44 | self.alphas_ = None 45 | self.coefs_ = None 46 | self.rankings_ = None 47 | 48 | def fit(self, X, y, feature_labels, estimator_params=None): 49 | """Computes the Lasso path using Sklearn's lasso_path method. 50 | 51 | Parameters 52 | ---------- 53 | X : array-like, shape (n_samples, n_features) 54 | Training data (the independent variables). 55 | 56 | y : array-like, shape (n_samples, n_outputs) 57 | Training data (the output/target values). 58 | 59 | feature_labels : array-like, shape (n_features) 60 | Labels for each of the features in X. 61 | 62 | estimator_params : dict, optional 63 | The parameters to pass to Sklearn's Lasso estimator. 64 | 65 | Returns 66 | ------- 67 | self 68 | """ 69 | 70 | self._reset() 71 | if estimator_params is None: 72 | estimator_params = {} 73 | self.feature_labels_ = feature_labels 74 | 75 | alphas, coefs, _ = lasso_path(X, y, **estimator_params) 76 | self.alphas_ = alphas.copy() 77 | self.coefs_ = coefs.copy() 78 | 79 | # Rank the features in X by order of importance. This ranking is based 80 | # on how early a given features enter the regression (the earlier a 81 | # feature enters the regression, the MORE important it is). 82 | feature_rankings = [[] for _ in range(X.shape[1])] 83 | for target_coef_paths in self.coefs_: 84 | for i, feature_path in enumerate(target_coef_paths): 85 | entrance_step = 1 86 | for val_at_step in feature_path: 87 | if val_at_step == 0: 88 | entrance_step += 1 89 | else: 90 | break 91 | feature_rankings[i].append(entrance_step) 92 | self.rankings_ = np.array([np.mean(ranks) for ranks in feature_rankings]) 93 | return self 94 | 95 | def get_ranked_features(self): 96 | if self.rankings_ is None: 97 | raise Exception("No lasso path has been fit yet!") 98 | 99 | rank_idxs = np.argsort(self.rankings_) 100 | return [self.feature_labels_[i] for i in rank_idxs] 101 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | gym==0.18.0 2 | 3 | Keras==2.3.0 4 | 5 | matplotlib==3.1.1 6 | 7 | numpy==1.17.2 8 | 9 | pandas==0.25.2 10 | 11 | paramiko==2.7.2 12 | 13 | PyMySQL==0.10.1 14 | 15 | pyparsing==2.4.2 16 | 17 | python-dateutil==2.8.0 18 | 19 | tensorflow==1.15.0 20 | 21 | sklearn==0.0 22 | 23 | DBUtils==2.0.1 24 | -------------------------------------------------------------------------------- /run_job.py: -------------------------------------------------------------------------------- 1 | import threading 2 | import time 3 | import queue 4 | import pymysql 5 | from configs import parse_args 6 | from dbutils.pooled_db import PooledDB 7 | 8 | args = parse_args() 9 | lock = threading.Lock() 10 | total_lat = 0 11 | error_query_num = 0 12 | 13 | POOL = None 14 | 15 | 16 | # 把任务放入队列中 17 | class Producer(threading.Thread): 18 | def __init__(self, name, queue, workload): 19 | self.__name = name 20 | self.__queue = queue 21 | self.workload = workload 22 | super(Producer, self).__init__() 23 | 24 | def run(self): 25 | for index, query in enumerate(self.workload): 26 | self.__queue.put(str(index) + "~#~" + query) 27 | 28 | 29 | # 线程处理任务 30 | class Consumer(threading.Thread): 31 | def __init__(self, name, queue): 32 | self.__name = name 33 | self.__queue = queue 34 | super(Consumer, self).__init__() 35 | 36 | def run(self): 37 | while not self.__queue.empty(): 38 | query = self.__queue.get() 39 | try: 40 | consumer_process(query) 41 | finally: 42 | self.__queue.task_done() 43 | 44 | 45 | def consumer_process(task_key): 46 | query = task_key.split('~#~')[1] 47 | if query: 48 | 49 | start = time.time() 50 | result = mysql_query(query) 51 | end = time.time() 52 | interval = end - start 53 | 54 | if result: 55 | lock.acquire() 56 | global total_lat 57 | total_lat += interval 58 | lock.release() 59 | 60 | else: 61 | global error_query_num 62 | lock.acquire() 63 | error_query_num += 1 64 | lock.release() 65 | 66 | 67 | def startConsumer(thread_num, queue): 68 | t_consumer = [] 69 | for i in range(thread_num): 70 | c = Consumer(i, queue) 71 | c.setDaemon(True) 72 | c.start() 73 | t_consumer.append(c) 74 | return t_consumer 75 | 76 | 77 | def run_job(thread_num=1, workload=[], resfile="../output.res"): 78 | global total_lat 79 | total_lat = 0 80 | global error_query_num 81 | error_query_num = 0 82 | workload_len = len(workload) 83 | 84 | global POOL 85 | POOL = PooledDB( 86 | creator=pymysql, # 使用链接数据库的模块 87 | maxconnections=thread_num, # 连接池允许的最大连接数，0和None表示不限制连接数 88 | mincached=0, # 初始化时，链接池中至少创建的空闲的链接，0表示不创建 89 | maxcached=0, # 链接池中最多闲置的链接，0和None不限制 90 | maxshared=0, 91 | blocking=True, # 连接池中如果没有可用连接后，是否阻塞等待。True，等待；False，不等待然后报错 92 | maxusage=None, # 一个链接最多被重复使用的次数，None表示无限制 93 | setsession=[], # 开始会话前执行的命令列表。 94 | ping=0, 95 | # ping MySQL服务端，检查是否服务可用。 96 | host=args["host"], 97 | port=int(args["port"]), 98 | user=args["user"], 99 | password=args["password"], 100 | database=args["database"], 101 | charset='utf8' 102 | ) 103 | 104 | main_queue = queue.Queue(maxsize=0) 105 | p = Producer("Producer Query", main_queue, workload) 106 | p.setDaemon(True) 107 | p.start() 108 | startConsumer(thread_num, main_queue) 109 | # 确保所有的任务都生成 110 | p.join() 111 | start = time.time() 112 | print("run_job开始运行,线程数：", thread_num) 113 | # 等待处理完所有任务 114 | main_queue.join() 115 | POOL.close() 116 | run_time = round(time.time() - start, 1) 117 | run_query_num = workload_len - error_query_num 118 | if run_query_num == 0: 119 | avg_lat = 0 120 | avg_qps = 0 121 | else: 122 | avg_lat = total_lat / run_query_num 123 | avg_qps = run_query_num / run_time 124 | text = "\navg_qps(queries/s): \t{}\navg_lat(s): \t{}\n".format(round(avg_qps, 4), round(avg_lat, 4)) 125 | with open(resfile, "w+") as f: 126 | f.write(text) 127 | f.close() 128 | print("run_job运行结束\n脚本总耗时:{}秒,sql执行总耗时:{}秒\n共有{}条数据，执行成功{}条\n{}".format(str(run_time), str(total_lat), 129 | str(workload_len), 130 | str(run_query_num), 131 | text)) 132 | 133 | return round(avg_qps, 4), round(avg_lat, 4) 134 | 135 | 136 | def mysql_query(sql: str) -> bool: 137 | try: 138 | global POOL 139 | conn = POOL.connection() 140 | cursor = conn.cursor() 141 | cursor.execute(sql) 142 | cursor.close() 143 | conn.commit() 144 | return True 145 | except Exception as error: 146 | print("mysql execute: " + str(error)) 147 | return False 148 | -------------------------------------------------------------------------------- /run_workload.py: -------------------------------------------------------------------------------- 1 | from configs import parse_args 2 | from get_workload_from_file import get_workload_from_file 3 | from run_job import run_job 4 | import time 5 | 6 | if __name__ == "__main__": 7 | 8 | argus = parse_args() 9 | 10 | # prepare_training_workloads 11 | workload = get_workload_from_file(argus["workload_file_path"]) 12 | 13 | file_path = 'training-results/res_no_change-' + str(int(time.time())) 14 | for i in range(0, 3): 15 | pfs = open(file_path, 'a') 16 | avg_qps, avg_lat = run_job(int(argus["thread_num"]), workload) 17 | pfs.write("%d\t%s\t%s\n" % (i, avg_qps, avg_lat)) 18 | pfs.close() 19 | 20 | -------------------------------------------------------------------------------- /saved_model_weights-141threads/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights-141threads/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights-141threads/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights-141threads/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights-2021-07-22-2/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights-2021-07-22-2/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights-2021-07-22-2/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights-2021-07-22-2/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights-2021-07-22/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights-2021-07-22/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights-2021-07-22/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights-2021-07-22/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights/rewards.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights/rewards.txt -------------------------------------------------------------------------------- /saved_model_weights_2021-07-07/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021-07-07/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021-07-07/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021-07-07/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021-07-08/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021-07-08/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021-07-08/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021-07-08/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021-07-09/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021-07-09/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021-07-09/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021-07-09/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021_07_02_23_39_12/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021_07_02_23_39_12/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021_07_02_23_39_12/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021_07_02_23_39_12/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021_07_11/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021_07_11/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021_07_11/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021_07_11/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021_07_11_2/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021_07_11_2/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_2021_07_11_2/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_2021_07_11_2/critic_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_tmp/actor_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_tmp/actor_weights.h5 -------------------------------------------------------------------------------- /saved_model_weights_tmp/critic_weights.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/XuanheZhou/qtune-mysql/741a42b1c4582ecce8c1cb0ee2aa6db052125af0/saved_model_weights_tmp/critic_weights.h5 -------------------------------------------------------------------------------- /scripts/output.res: -------------------------------------------------------------------------------- 1 | 2 | avg_qps(queries/s): 753.9394 3 | avg_lat(s): 0.1227 4 | -------------------------------------------------------------------------------- /scripts/run_job.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | # run_job.sh selectedList.txt queries_dir output 3 | 4 | avg_lat=0 5 | avg_tps=0 6 | count=0 7 | 8 | printf "query\tlat(ms)\n" > $7 9 | 10 | tmp_fifofile="/tmp/$$.fifo" 11 | mkfifo $tmp_fifofile 12 | exec 6<>$tmp_fifofile 13 | rm $tmp_fifofile 14 | 15 | thread_num=$8 16 | 17 | rm "/tmp/avg_lat_pipef.txt" 18 | rm "/tmp/avg_tps_pipef.txt" 19 | rm "/tmp/count_pipef.txt" 20 | 21 | for ((i=0;i<${thread_num};i++));do 22 | echo 23 | done >&6 24 | 25 | for file in $6/*; 26 | do 27 | read -u6 28 | { 29 | tmp=$(mysql -h$1 -p$2 -u$3 -p$4 $5 < $file | tail -n 1 ) 30 | query=`echo $tmp | awk '{print $1}'` 31 | lat=`echo $tmp | awk '{print $2}'` 32 | mysql -h$1 -p$2 -u$3 -p$4 $5 -e"\q" 33 | tps=$(echo "scale=4; 60000 / $lat" | bc) 34 | 35 | echo "scale=4; $lat / 1000" | bc >> "/tmp/avg_lat_pipef.txt" 36 | echo "scale=4; $tps" | bc >> "/tmp/avg_tps_pipef.txt" 37 | echo "1" | bc >> "/tmp/count_pipef.txt" 38 | printf "$query\t$lat\n" >> $7 39 | echo >&6 40 | } & 41 | done 42 | wait 43 | exec 6>&- 44 | 45 | avg_lat=$(echo $(echo -n `cat /tmp/avg_lat_pipef.txt | awk '{print $1}'`| tr ' ' '+')|bc) 46 | avg_tps=$(echo $(echo -n `cat /tmp/avg_tps_pipef.txt | awk '{print $1}'`| tr ' ' '+')|bc) 47 | count=$(echo $(echo -n `cat /tmp/count_pipef.txt | awk '{print $1}'`| tr ' ' '+')|bc) 48 | 49 | rm "/tmp/avg_lat_pipef.txt" 50 | rm "/tmp/avg_tps_pipef.txt" 51 | rm "/tmp/count_pipef.txt" 52 | 53 | avg_lat=`echo "scale=4; $avg_lat/$count" | bc` 54 | avg_tps=`echo "scale=4; $avg_tps/$count" | bc` 55 | 56 | printf "\navg_tps(txn/min): \t%5.4f\navg_lat(ms): \t%5.4f\n" $avg_tps $avg_lat >> $7 57 | -------------------------------------------------------------------------------- /sql2resource.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import pandas 3 | import json 4 | import os 5 | import pymysql 6 | from keras.models import Sequential, Model 7 | from keras.layers import Dense, Dropout, Input 8 | from configs import predictor_output_dim 9 | 10 | query_types = ["insert", "delete", "update", "select"] 11 | 12 | 13 | # base prediction model 14 | def baseline_model(num_feature=len(query_types)): 15 | # create model 16 | model = Sequential() 17 | model.add(Dense(10, input_dim=num_feature, kernel_initializer='normal', activation='relu')) 18 | model.add(Dense(10, kernel_initializer='normal', activation='relu')) 19 | model.add(Dense(predictor_output_dim, kernel_initializer='normal')) 20 | # Compile model 21 | model.compile(loss='mean_squared_error', optimizer='adam') 22 | 23 | return model 24 | 25 | 26 | class SqlParser: 27 | ########################################################################### 28 | # DML: select delete insert update 0 1 2 3 29 | # select {select_expr} 30 | # (The first is default) 31 | # [ALL | DISTINCT | DISTINCTROW] 32 | # [0 | HIGH_PRIORITY], faster than update, with table-level lock 33 | # [0 | STRAIGHT_JOIN], 34 | # [0 | SQL_SMALL_RESULT | SQL_BIG_RESULT] 35 | # [0 | SQL_BUFFER_RESULT] 36 | # [SQL_CACHE | SQL_NO_CACHE] 37 | # [SQL_CALC_FOUND_ROWS] 38 | 39 | # OLTP (workload2vector) 40 | # select*w1 + sum(modifiers)*w2 + num({select_expr})*wl3 # 0.7 0.1 0.2 41 | # from [table] 42 | # [WHERE where_condition] range join 43 | 44 | # OLTP (sql2vector) 45 | # cost-vector: [Aggregate, Nested Loop, Index Scan, Hash_Join] 46 | 47 | # Keywords 48 | # [GROUP BY {col_name | expr | position}] 49 | # [ASC | DESC], ...[WITH ROLLUP] 50 | # [HAVING where_condition] 51 | # [ORDER BY {col_name | expr | position}] 52 | # [ASC | DESC], ... 53 | 54 | # sum(group_table_scale(having)*wi) + order_cost*wi 55 | ########################################################################### 56 | 57 | def __init__(self, argus): 58 | 59 | self.resfile = os.path.join("scripts/") + "output.res" 60 | self.argus = argus 61 | self.conn = self.mysql_conn() 62 | self.tables = self.get_database_tables() 63 | self.query_encoding_map = {} 64 | ########### Convert from the sql statement to the sql vector 65 | # directly read vector from a file (so a python2 script needs to run first!) 66 | # sql_type * (num_events, C, aggregation, in-mem) 67 | ############################################################################################################################# 68 | 69 | # query encoding features 70 | self.op_weight = {'oltp_point_select': 1, 'select_random_ranges': 2, 'oltp_delete': 3, 71 | 'oltp_insert': 4, 'bulk_insert': 5, 'oltp_update_index': 6, 72 | 'oltp_update_non_index': 7, } 73 | self.num_event = int(argus['num_event']) 74 | self.C = [10000] 75 | self.group_cost = 0 76 | self.in_mem = 0 77 | self.predict_sql_resource_value = None 78 | self.estimator = baseline_model(len(query_types) + len(self.tables)) 79 | # Prepare Data 80 | fs = open("training-data/trainData_sql.txt", 'r') 81 | df = pandas.read_csv(fs, sep=' ', header=None) 82 | lt_sql = df.values 83 | # seperate into input X and output Y 84 | sql_op = lt_sql[:, 0] 85 | sql_X = lt_sql[:, 1:5] # op_type events table_size 86 | sql_Y = lt_sql[:, 5:] 87 | 88 | def query_encoding(self, query): 89 | 90 | if not query: 91 | print("query is empty") 92 | return [] 93 | 94 | if self.query_encoding_map.get(str(query), None): 95 | return self.query_encoding_map[str(query)] 96 | 97 | result = [0 for i in range(len(self.tables) + len(query_types))] 98 | # [0, 0, 0, 0, X, X, X..........] 99 | query_split_list = query.lower().split(" ") 100 | 101 | for index, query_type in enumerate(query_types): 102 | if query_type in query_split_list: 103 | result[index] = 1 104 | 105 | query_split_list = query.replace(",", "").split(" ") 106 | 107 | explain_format_fetchall = self.mysql_query("EXPLAIN FORMAT=JSON {};".format(query)) 108 | if not explain_format_fetchall: 109 | print("explain_format_fetchall is empty, query: {}".format(query)) 110 | return [] 111 | explain_format = json.loads(explain_format_fetchall[0][0]) 112 | explain_format_tables_list = self.get_explain_format_tables_list([], explain_format.get("query_block"), "table") 113 | for explain_format_table in explain_format_tables_list: 114 | explain_format_table_name = explain_format_table["table_name"] 115 | index = query_split_list.index(explain_format_table_name) 116 | if query_split_list[index - 1].lower() == "as": 117 | explain_format_table_name = query_split_list[index - 2] 118 | else: 119 | explain_format_table_name = query_split_list[index - 1] 120 | 121 | for index, table_name in enumerate(self.tables): 122 | if explain_format_table_name == table_name: 123 | result[index + len(query_types)] = float(explain_format_table["cost_info"]["prefix_cost"]) 124 | continue 125 | self.query_encoding_map[str(query)] = result 126 | return result 127 | 128 | def predict_sql_resource(self, workload=[]): 129 | # Predict sql convert 130 | # inner_metric_change np.array 131 | if self.predict_sql_resource_value is None: 132 | print("predict_sql_resource_value is None") 133 | exit() 134 | return self.predict_sql_resource_value 135 | # return self.estimator.predict(self.get_workload_encoding( 136 | # workload)) # input : np.array([[...]]) (sq_type, num_events, C, aggregation, in-mem) 137 | # # output : np.array([[...]]) 138 | 139 | def update(self): 140 | pass 141 | 142 | def mysql_conn(self): 143 | conn = pymysql.connect( 144 | host=self.argus["host"], 145 | user=self.argus["user"], 146 | passwd=self.argus["password"], 147 | port=int(self.argus["port"]), 148 | connect_timeout=30, 149 | charset='utf8') 150 | conn.select_db(self.argus["database"]) 151 | return conn 152 | 153 | def close_mysql_conn(self): 154 | try: 155 | self.conn.close() 156 | except Exception as error: 157 | print("close mysqlconn: " + str(error)) 158 | 159 | def mysql_query(self, sql): 160 | try: 161 | cursor = self.conn.cursor() 162 | count = cursor.execute(sql) 163 | if count == 0: 164 | result = 0 165 | else: 166 | result = cursor.fetchall() 167 | cursor.close() 168 | return result 169 | except Exception as error: 170 | print("mysql execute: " + str(error)) 171 | return None 172 | 173 | def get_database_tables(self): 174 | # get all tables 175 | tables_fetchall = self.mysql_query( 176 | "select table_name from information_schema.tables where table_schema='{}';".format(self.argus["database"])) 177 | tables = [] 178 | if not tables_fetchall: 179 | print("tables was not found") 180 | return 181 | for table in tables_fetchall: 182 | if table and table[0]: 183 | tables.append(table[0]) 184 | print("get all tables success") 185 | return tables 186 | 187 | def get_explain_format_tables_list(self, values_list, json, key): 188 | if isinstance(json, dict): 189 | for item, values in json.items(): 190 | if str(item) == str(key): 191 | values_list.append(json.get(item)) 192 | if isinstance(values, dict): 193 | self.get_explain_format_tables_list(values_list, values, key=key) 194 | if isinstance(values, list): 195 | self.get_explain_format_tables_list(values_list, values, key=key) 196 | else: 197 | pass 198 | elif isinstance(json, list): 199 | for data in json: 200 | if isinstance(data, dict): 201 | self.get_explain_format_tables_list(values_list, data, key) 202 | else: 203 | return values_list 204 | return values_list 205 | 206 | def get_workload_encoding(self, workload): 207 | queries_encoding = [] 208 | for query in workload: 209 | queries_encoding.append(self.query_encoding(query)) 210 | 211 | # [0, 0, 0, 0, X, X, X..........] 212 | workload_encoding = np.array([0 for i in range(len(self.tables) + len(query_types))]) 213 | for query_encoding in queries_encoding: 214 | workload_encoding = workload_encoding + np.array(query_encoding) 215 | 216 | for i in range(len(query_types)): 217 | if workload_encoding[i] > 0: 218 | workload_encoding[i] = 1 219 | 220 | return workload_encoding.reshape(1, len(workload_encoding)) 221 | -------------------------------------------------------------------------------- /training-data/trainData_sql.txt: -------------------------------------------------------------------------------- 1 | 0 10 10000 0 0 0 0 0 0 0 0 0 0 0 18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 2 | -------------------------------------------------------------------------------- /workload_file_example.txt: -------------------------------------------------------------------------------- 1 | SELECT MIN(mi_idx.info) AS rating, MIN(t.title) AS movie_title FROM info_type AS it, keyword AS k, movie_info_idx AS mi_idx, movie_keyword AS mk, title AS t WHERE it.info ='rating' AND k.keyword like '%sequel%' AND mi_idx.info > '9.0' AND t.production_year > 2010 AND t.id = mi_idx.movie_id AND t.id = mk.movie_id AND mk.movie_id = mi_idx.movie_id AND k.id = mk.keyword_id AND it.id = mi_idx.info_type_id; 2 | 3 | SELECT MIN(mi_idx.info) AS rating, MIN(t.title) AS movie_title FROM info_type AS it, keyword AS k, movie_info_idx AS mi_idx, movie_keyword AS mk, title AS t WHERE it.info ='rating' AND k.keyword like '%sequel%' AND mi_idx.info > '5.0' AND t.production_year > 2005 AND t.id = mi_idx.movie_id AND t.id = mk.movie_id AND mk.movie_id = mi_idx.movie_id AND k.id = mk.keyword_id AND it.id = mi_idx.info_type_id; 4 | 5 | SELECT MIN(mi_idx.info) AS rating, MIN(t.title) AS movie_title FROM info_type AS it, keyword AS k, movie_info_idx AS mi_idx, movie_keyword AS mk, title AS t WHERE it.info ='rating' AND k.keyword like '%sequel%' AND mi_idx.info > '2.0' AND t.production_year > 1990 AND t.id = mi_idx.movie_id AND t.id = mk.movie_id AND mk.movie_id = mi_idx.movie_id AND k.id = mk.keyword_id AND it.id = mi_idx.info_type_id; 6 | 7 | -------------------------------------------------------------------------------- /配置文档.md: -------------------------------------------------------------------------------- 1 | ### 功能 2 | 3 | [✓] 基于负载+状态特征自动调参（environment.py） 4 | 5 | [✓] 解决硬编码问题（database、benchmark、knob类别|范围|类型等、系统状态指标、prediction model、模型尺寸等） 6 | 7 | [✓] 提供预训练数据和模型（model.py） 8 | 9 | [✓] 添加OLAP/OLTP测试集（JOB、Sysbench、TPC-H） 10 | 11 | [✓] 绘制训练曲线（draw.py），包括随机探索、模型预测的表现（latency/throughput） 12 | 13 | [✓] 停机问题 14 | 15 | 设置config.ini文件中的stopping_throughput_improvement_percentage参数。 16 | 17 | [todo] 同时支持重启/非重启参数 18 | 19 | [todo] 支持基于模板的在线参数调优；端到端训练调参模型和负载编码模型 20 | 21 | [todo] 集合现有规则预调参 22 | 23 | 24 | ### 搭建步骤 25 | 26 | 1. 安装v 5.7 MySQL数据库 27 | 28 | * 给mysql的root用户开启远程访问权限参考：[mysql给root开启远程访问权限](https://www.cnblogs.com/goxcheer/p/8797377.html) 29 | 30 | * add a new line `sql_mode=NO_ENGINE_SUBSTITUTION` to `my.cnf` or `my.ini`, and restart mysqld 31 | 32 | 2. 在数据库的服务器上，上传标准测试集的代码，并进行安装： 33 | * Sysbench：https://blog.csdn.net/cxin917/article/details/81557453（数据库服务器端安装） 34 | ``` 35 | sysbench oltp_read_only --db-driver=mysql --tables=4 --table-size=4000000 --threads=141 --events=0 --mysql-host=172.27.58.68 --mysql-user=root --mysql-password=dbmind2020 --mysql-port=3306 --mysql-db='sysbench' --time=150 --range-size=10 --mysql-storage-engine=innodb --rand-type=uniform --report-interval=10 36 | ``` 37 | * JOB：https://blog.csdn.net/cxin917/article/details/81557453 (imdb数据集如果下载过慢，可以通过云盘提供) 38 | * 开启mysqllog，获取workload。 39 | * 关闭mysqllog 40 | 3. 将调参代码放在能够连接数据库的一台服务器上，配置python3.6依赖环境。安装依赖包 41 | ``` 42 | pip3 install -r requirements.txt 43 | ``` 44 | 4. 修改参数: 45 | 将config_example.ini复制，新文件名为config.ini 46 | ` 47 | cp config_example.ini config.ini 48 | ` 49 | 在config.ini修改各项参数 50 | workload_file_path将从指定的文件中获取，文件内容可参考workload_file_example.txt文件。 51 | 52 | 5. 运行代码： 53 | 54 | ``` 55 | python3 main.py 56 | ``` 57 | 58 | 6. 手动绘制training.png，用于观察训练效果： 59 | 60 | ``` 61 | python3 draw.py res_predict-1623852012,res_random-1623852012 latency 62 | ``` 63 | 64 | 65 | ### 报错解决 66 | 67 | **Q:** “mysql cannot connect from remote host” 68 | 69 | **A:** https://devanswers.co/cant-connect-mysql-server-remotely/#:~:text=You%20may%20need%20to%20comment,the%20MySQL%20config%20file%20mysqld.&text=The%20above%20line%20is%20telling,remote%20connections%20to%20that%20IP. 70 | 71 | **Q:** JOB queries take very long time to run 72 | 73 | **A:** 删除所有外键并创建索引（修改fkindexes.sql）；将脚本改成并行执行；选择部分sql执行 74 | 75 | **Q:** “ImportError: attempted relative import with no known parent package” 76 | 77 | **A:** **https://stackoverflow.com/questions/16981921/relative-imports-in-python-3 78 | 79 | **Q:** pd.read_csv: “ValueError: cannot convert float NaN to integer“ 80 | 81 | **A:** 1. , sep = "\t"; 2. encoding style (utf-8) 82 | --------------------------------------------------------------------------------