restore CDBTune

server/analysis/ddpg/ddpg.py

@@ -3,7 +3,7 @@
 #
 # Copyright (c) 2017-18, Carnegie Mellon University Database Group
 #
-# from: https://github.com/KqSMea8/CDBTune
+# from: https://github.com/KqSMea8/use_default
 # Zhang, Ji, et al. "An end-to-end automatic cloud database tuning system using
 # deep reinforcement learning." Proceedings of the 2019 International Conference
 # on Management of Data. ACM, 2019
@@ -23,8 +23,22 @@ LOG = get_analysis_logger(__name__)
 
 class Actor(nn.Module):
 
-    def __init__(self, n_states, n_actions, hidden_sizes):
+    def __init__(self, n_states, n_actions, hidden_sizes, use_default):
         super(Actor, self).__init__()
+        if use_default:
+            self.layers = nn.Sequential(
+                nn.Linear(n_states, 128),
+                nn.LeakyReLU(negative_slope=0.2),
+                nn.BatchNorm1d(hidden_sizes[0]),
+                nn.Linear(128, 128),
+                nn.Tanh(),
+                nn.Dropout(0.3),
+                nn.Linear(128, 128),
+                nn.Tanh(),
+                nn.Linear(128, 64),
+                nn.Linear(64, n_actions)
+            )
+        else:
             self.layers = nn.Sequential(
                 nn.Linear(n_states, hidden_sizes[0]),
                 nn.LeakyReLU(negative_slope=0.2),
@@ -58,11 +72,26 @@ class Actor(nn.Module):
 
 class Critic(nn.Module):
 
-    def __init__(self, n_states, n_actions, hidden_sizes):
+    def __init__(self, n_states, n_actions, hidden_sizes, use_default):
         super(Critic, self).__init__()
+        self.act = nn.Tanh()
+        if use_default:
+            self.state_input = nn.Linear(n_states, 128)
+            self.action_input = nn.Linear(n_actions, 128)
+            self.layers = nn.Sequential(
+                nn.Linear(256, 256),
+                nn.LeakyReLU(negative_slope=0.2),
+                nn.BatchNorm1d(256),
+                nn.Linear(256, 256),
+                nn.Linear(256, 64),
+                nn.Tanh(),
+                nn.Dropout(0.3),
+                nn.BatchNorm1d(64),
+                nn.Linear(64, 1)
+            )
+        else:
             self.state_input = nn.Linear(n_states, hidden_sizes[0])
             self.action_input = nn.Linear(n_actions, hidden_sizes[0])
-        self.act = nn.Tanh()
             self.layers = nn.Sequential(
                 nn.Linear(hidden_sizes[0] * 2, hidden_sizes[1]),
                 nn.LeakyReLU(negative_slope=0.2),
@@ -72,7 +101,7 @@ class Critic(nn.Module):
                 nn.Tanh(),
                 nn.Dropout(0.3),
                 nn.BatchNorm1d(hidden_sizes[2]),
-            nn.Linear(hidden_sizes[2], 1),
+                nn.Linear(hidden_sizes[2], 1)
             )
         self._init_weights()
 
@@ -101,7 +130,8 @@ class DDPG(object):
 
     def __init__(self, n_states, n_actions, model_name='', alr=0.001, clr=0.001,
                  gamma=0.9, batch_size=32, tau=0.002, shift=0, memory_size=100000,
-                 a_hidden_sizes=[128, 128, 64], c_hidden_sizes=[128, 256, 64]):
+                 a_hidden_sizes=[128, 128, 64], c_hidden_sizes=[128, 256, 64],
+                 use_default=False):
         self.n_states = n_states
         self.n_actions = n_actions
         self.alr = alr
@@ -113,6 +143,7 @@ class DDPG(object):
         self.a_hidden_sizes = a_hidden_sizes
         self.c_hidden_sizes = c_hidden_sizes
         self.shift = shift
+        self.use_default = use_default
 
         self._build_network()
 
@@ -124,10 +155,12 @@ class DDPG(object):
         return Variable(torch.FloatTensor(x))
 
     def _build_network(self):
-        self.actor = Actor(self.n_states, self.n_actions, self.a_hidden_sizes)
+        self.actor = Actor(self.n_states, self.n_actions, self.a_hidden_sizes, self.use_default)
-        self.target_actor = Actor(self.n_states, self.n_actions, self.a_hidden_sizes)
+        self.target_actor = Actor(self.n_states, self.n_actions, self.a_hidden_sizes,
-        self.critic = Critic(self.n_states, self.n_actions, self.c_hidden_sizes)
+                                  self.use_default)
-        self.target_critic = Critic(self.n_states, self.n_actions, self.c_hidden_sizes)
+        self.critic = Critic(self.n_states, self.n_actions, self.c_hidden_sizes, self.use_default)
+        self.target_critic = Critic(self.n_states, self.n_actions, self.c_hidden_sizes,
+                                    self.use_default)
 
         # Copy actor's parameters
         self._update_target(self.target_actor, self.actor, tau=1.0)

server/website/website/settings/constants.py

@@ -52,11 +52,11 @@ DEFAULT_LEARNING_RATE = 0.01
 #  a small bias when using training data points as starting points.
 GPR_EPS = 0.001
 
-DEFAULT_RIDGE = 1.0
+DEFAULT_RIDGE = 0.01
 
 DEFAULT_EPSILON = 1e-6
 
-DEFAULT_SIGMA_MULTIPLIER = 1.0
+DEFAULT_SIGMA_MULTIPLIER = 3.0
 
 DEFAULT_MU_MULTIPLIER = 1.0
 
@@ -84,6 +84,13 @@ DNN_DEBUG = True
 DNN_DEBUG_INTERVAL = 100
 
 # ---DDPG CONSTRAINTS CONSTANTS---
+
+#  Use a simple reward
+DDPG_SIMPLE_REWARD = True
+
+#  The weight of future rewards in Q value
+DDPG_GAMMA = 0.0
+
 #  Batch size in DDPG model
 DDPG_BATCH_SIZE = 32
 
@@ -101,3 +108,14 @@ ACTOR_HIDDEN_SIZES = [128, 128, 64]
 
 #  The number of hidden units in each layer of the critic MLP
 CRITIC_HIDDEN_SIZES = [64, 128, 64]
+
+#  Use the same setting from the CDBTune paper
+USE_DEFAULT = True
+#  Overwrite the DDPG settings if using CDBTune
+if USE_DEFAULT:
+    DDPG_SIMPLE_REWARD = False
+    DDPG_GAMMA = 0.99
+    DDPG_BATCH_SIZE = 32
+    ACTOR_LEARNING_RATE = 0.001
+    CRITIC_LEARNING_RATE = 0.001
+    UPDATE_EPOCHS = 1

server/website/website/tasks/async_tasks.py

@@ -36,6 +36,7 @@ from website.settings import (USE_GPFLOW, DEFAULT_LENGTH_SCALE, DEFAULT_MAGNITUD
                               DEFAULT_EPSILON, MAX_ITER, GPR_EPS,
                               DEFAULT_SIGMA_MULTIPLIER, DEFAULT_MU_MULTIPLIER,
                               DEFAULT_UCB_SCALE, HP_LEARNING_RATE, HP_MAX_ITER,
+                              DDPG_SIMPLE_REWARD, DDPG_GAMMA, USE_DEFAULT,
                               DDPG_BATCH_SIZE, ACTOR_LEARNING_RATE,
                               CRITIC_LEARNING_RATE, UPDATE_EPOCHS,
                               ACTOR_HIDDEN_SIZES, CRITIC_HIDDEN_SIZES,
@@ -285,18 +286,25 @@ def train_ddpg(result_id):
     result = Result.objects.get(pk=result_id)
     session = Result.objects.get(pk=result_id).session
     session_results = Result.objects.filter(session=session,
-                                            creation_time__lte=result.creation_time)
+                                            creation_time__lt=result.creation_time)
     result_info = {}
     result_info['newest_result_id'] = result_id
 
-    # Extract data from result
+    # Extract data from result and previous results
     result = Result.objects.filter(pk=result_id)
+    if len(session_results) == 0:
+        base_result_id = result_id
+        prev_result_id = result_id
+    else:
         base_result_id = session_results[0].pk
+        prev_result_id = session_results[len(session_results)-1].pk
     base_result = Result.objects.filter(pk=base_result_id)
+    prev_result = Result.objects.filter(pk=prev_result_id)
 
     agg_data = DataUtil.aggregate_data(result)
     metric_data = agg_data['y_matrix'].flatten()
     base_metric_data = (DataUtil.aggregate_data(base_result))['y_matrix'].flatten()
+    prev_metric_data = (DataUtil.aggregate_data(prev_result))['y_matrix'].flatten()
     metric_scalar = MinMaxScaler().fit(metric_data.reshape(1, -1))
     normalized_metric_data = metric_scalar.transform(metric_data.reshape(1, -1))[0]
 
@@ -323,21 +331,37 @@ def train_ddpg(result_id):
                                                            target_objective))
     objective = metric_data[target_obj_idx]
     base_objective = base_metric_data[target_obj_idx]
+    prev_objective = prev_metric_data[target_obj_idx]
     metric_meta = db.target_objectives.get_metric_metadata(
         result.session.dbms.pk, result.session.target_objective)
 
     # Calculate the reward
+    if DDPG_SIMPLE_REWARD:
         objective = objective / base_objective
         if metric_meta[target_objective].improvement == '(less is better)':
             reward = -objective
         else:
             reward = objective
+    else:
+        if metric_meta[target_objective].improvement == '(less is better)':
+            if objective - base_objective <= 0:  # positive reward
+                reward = (np.square((2 * base_objective - objective) / base_objective) - 1)\
+                    * abs(2 * prev_objective - objective) / prev_objective
+            else:  # negative reward
+                reward = -(np.square(objective / base_objective) - 1) * objective / prev_objective
+        else:
+            if objective - base_objective > 0:  # positive reward
+                reward = (np.square(objective / base_objective) - 1) * objective / prev_objective
+            else:  # negative reward
+                reward = -(np.square((2 * base_objective - objective) / base_objective) - 1)\
+                    * abs(2 * prev_objective - objective) / prev_objective
     LOG.info('reward: %f', reward)
 
     # Update ddpg
     ddpg = DDPG(n_actions=knob_num, n_states=metric_num, alr=ACTOR_LEARNING_RATE,
-                clr=CRITIC_LEARNING_RATE, gamma=0, batch_size=DDPG_BATCH_SIZE,
+                clr=CRITIC_LEARNING_RATE, gamma=DDPG_GAMMA, batch_size=DDPG_BATCH_SIZE,
-                a_hidden_sizes=ACTOR_HIDDEN_SIZES, c_hidden_sizes=CRITIC_HIDDEN_SIZES)
+                a_hidden_sizes=ACTOR_HIDDEN_SIZES, c_hidden_sizes=CRITIC_HIDDEN_SIZES,
+                use_default=USE_DEFAULT)
     if session.ddpg_actor_model and session.ddpg_critic_model:
         ddpg.set_model(session.ddpg_actor_model, session.ddpg_critic_model)
     if session.ddpg_reply_memory:
@@ -368,7 +392,7 @@ def configuration_recommendation_ddpg(result_info):  # pylint: disable=invalid-n
     metric_num = len(metric_data)
 
     ddpg = DDPG(n_actions=knob_num, n_states=metric_num, a_hidden_sizes=ACTOR_HIDDEN_SIZES,
-                c_hidden_sizes=CRITIC_HIDDEN_SIZES)
+                c_hidden_sizes=CRITIC_HIDDEN_SIZES, use_default=USE_DEFAULT)
     if session.ddpg_actor_model is not None and session.ddpg_critic_model is not None:
         ddpg.set_model(session.ddpg_actor_model, session.ddpg_critic_model)
     if session.ddpg_reply_memory is not None:
@@ -646,9 +670,8 @@ def configuration_recommendation(recommendation_input):
                           epsilon=DEFAULT_EPSILON,
                           max_iter=MAX_ITER,
                           sigma_multiplier=DEFAULT_SIGMA_MULTIPLIER,
-                          mu_multiplier=DEFAULT_MU_MULTIPLIER,
+                          mu_multiplier=DEFAULT_MU_MULTIPLIER)
-                          ridge=DEFAULT_RIDGE)
+            model.fit(X_scaled, y_scaled, X_min, X_max, ridge=DEFAULT_RIDGE)
-            model.fit(X_scaled, y_scaled, X_min, X_max)
             res = model.predict(X_samples, constraint_helper=constraint_helper)
 
     best_config_idx = np.argmin(res.minl.ravel())