public/ottertune
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

save ddpg model in database

Browse Source
This commit is contained in:
Dongsheng Yang
2019-09-26 14:44:28 -04:00
committed by Dana Van Aken
parent c8fbaf6e4b
commit a3fcf59f07
12 changed files with 683 additions and 736 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
server/website/website/migrations/0001_initial.py
View File

@@ -185,6 +185,9 @@ class Migration(migrations.Migration):
('id', models.AutoField(auto_created=True, primary_key=True, serialize=False, verbose_name='ID')),
('name', models.CharField(max_length=64, verbose_name=b'session name')),
('description', models.TextField(blank=True, null=True)),
('ddpg_actor_model', models.BinaryField(null=True, blank=True)),
('ddpg_critic_model', models.BinaryField(null=True, blank=True)),
('ddpg_reply_memory', models.BinaryField(null=True, blank=True)),
('creation_time', models.DateTimeField()),
('last_update', models.DateTimeField()),
('upload_code', models.CharField(max_length=30, unique=True)),

3
server/website/website/models.py
View File

@@ -187,6 +187,9 @@ class Session(BaseModel):
hardware = models.ForeignKey(Hardware)
algorithm = models.IntegerField(choices=AlgorithmType.choices(),
default=AlgorithmType.OTTERTUNE)
ddpg_actor_model = models.BinaryField(null=True, blank=True)
ddpg_critic_model = models.BinaryField(null=True, blank=True)
ddpg_reply_memory = models.BinaryField(null=True, blank=True)
project = models.ForeignKey(Project)
creation_time = models.DateTimeField()

2
server/website/website/parser/base.py
View File

@@ -353,7 +353,7 @@ class BaseParser(object, metaclass=ABCMeta):
def format_enum(self, enum_value, metadata):
enumvals = metadata.enumvals.split(',')
return enumvals[enum_value]
return enumvals[int(round(enum_value))]
def format_integer(self, int_value, metadata):
return int(round(int_value))

18
server/website/website/settings/constants.py
View File

@@ -35,7 +35,7 @@ MAX_TRAIN_SIZE = 7000
# Batch size in GPR model
BATCH_SIZE = 3000
# Threads for TensorFlow config
# Threads for TensorFlow config
NUM_THREADS = 4
# ---GRADIENT DESCENT CONSTANTS---
@@ -54,3 +54,19 @@ DEFAULT_EPSILON = 1e-6
DEFAULT_SIGMA_MULTIPLIER = 3.0
DEFAULT_MU_MULTIPLIER = 1.0
# ---CONSTRAINTS CONSTANTS---
# Batch size in DDPG model
DDPG_BATCH_SIZE = 32
# Learning rate of actor network
ACTOR_LEARNING_RATE = 0.001
# Learning rate of critic network
CRITIC_LEARNING_RATE = 0.001
# The impact of future reward on the decision
GAMMA = 0.1
# The changing rate of the target network
TAU = 0.002

2
server/website/website/tasks/__init__.py
View File

@@ -7,7 +7,7 @@ from .async_tasks import (aggregate_target_results,
configuration_recommendation,
map_workload,
train_ddpg,
run_ddpg)
configuration_recommendation_ddpg)
from .periodic_tasks import (run_background_tasks)

62
server/website/website/tasks/async_tasks.py
View File

@@ -5,14 +5,12 @@
#
import random
import queue
from os.path import dirname, abspath, join
import os
import numpy as np
from celery.task import task, Task
from celery.utils.log import get_task_logger
from djcelery.models import TaskMeta
from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import StandardScaler, MinMaxScaler
from analysis.ddpg.ddpg import DDPG
from analysis.gp import GPRNP
@@ -29,7 +27,10 @@ from website.settings import (DEFAULT_LENGTH_SCALE, DEFAULT_MAGNITUDE,
MAX_TRAIN_SIZE, BATCH_SIZE, NUM_THREADS,
DEFAULT_RIDGE, DEFAULT_LEARNING_RATE,
DEFAULT_EPSILON, MAX_ITER, GPR_EPS,
DEFAULT_SIGMA_MULTIPLIER, DEFAULT_MU_MULTIPLIER)
DEFAULT_SIGMA_MULTIPLIER, DEFAULT_MU_MULTIPLIER,
DDPG_BATCH_SIZE, ACTOR_LEARNING_RATE,
CRITIC_LEARNING_RATE, GAMMA, TAU)
from website.settings import INIT_FLIP_PROB, FLIP_PROB_DECAY
from website.types import VarType
@@ -235,10 +236,10 @@ def train_ddpg(result_id):
# Clean knob data
cleaned_agg_data = clean_knob_data(agg_data['X_matrix'], agg_data['X_columnlabels'], session)
agg_data['X_matrix'] = np.array(cleaned_agg_data[0]).flatten()
agg_data['X_columnlabels'] = np.array(cleaned_agg_data[1]).flatten()
knob_data = DataUtil.normalize_knob_data(agg_data['X_matrix'],
agg_data['X_columnlabels'], session)
knob_data = np.array(cleaned_agg_data[0])
knob_labels = np.array(cleaned_agg_data[1])
knob_bounds = np.vstack(DataUtil.get_knob_bounds(knob_labels.flatten(), session))
knob_data = MinMaxScaler().fit(knob_bounds).transform(knob_data)[0]
knob_num = len(knob_data)
metric_num = len(metric_data)
LOG.info('knob_num: %d, metric_num: %d', knob_num, metric_num)
@@ -276,26 +277,23 @@ def train_ddpg(result_id):
* (2 * prev_objective - objective) / prev_objective
# Update ddpg
project_root = dirname(dirname(dirname(abspath(__file__))))
saved_memory = join(project_root, 'checkpoint/reply_memory_' + session.project.name)
saved_model = join(project_root, 'checkpoint/ddpg_' + session.project.name)
ddpg = DDPG(n_actions=knob_num, n_states=metric_num)
if os.path.exists(saved_memory):
ddpg.replay_memory.load_memory(saved_memory)
ddpg.load_model(saved_model)
ddpg = DDPG(n_actions=knob_num, n_states=metric_num, alr=ACTOR_LEARNING_RATE,
clr=CRITIC_LEARNING_RATE, gamma=GAMMA, batch_size=DDPG_BATCH_SIZE, tau=TAU)
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:
ddpg.replay_memory.set(session.ddpg_reply_memory)
ddpg.add_sample(prev_metric_data, knob_data, reward, metric_data, False)
if len(ddpg.replay_memory) > 32:
ddpg.update()
checkpoint_dir = join(project_root, 'checkpoint')
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
ddpg.replay_memory.save(saved_memory)
ddpg.save_model(saved_model)
session.ddpg_actor_model, session.ddpg_critic_model = ddpg.get_model()
session.ddpg_reply_memory = ddpg.replay_memory.get()
session.save()
return result_info
@task(base=ConfigurationRecommendation, name='run_ddpg')
def run_ddpg(result_info):
@task(base=ConfigurationRecommendation, name='configuration_recommendation_ddpg')
def configuration_recommendation_ddpg(result_info): # pylint: disable=invalid-name
LOG.info('Use ddpg to recommend configuration')
result_id = result_info['newest_result_id']
result = Result.objects.filter(pk=result_id)
@@ -305,20 +303,20 @@ def run_ddpg(result_info):
cleaned_agg_data = clean_knob_data(agg_data['X_matrix'], agg_data['X_columnlabels'],
session)
knob_labels = np.array(cleaned_agg_data[1]).flatten()
knob_data = np.array(cleaned_agg_data[0]).flatten()
knob_num = len(knob_data)
knob_num = len(knob_labels)
metric_num = len(metric_data)
project_root = dirname(dirname(dirname(abspath(__file__))))
saved_memory = join(project_root, 'checkpoint/reply_memory_' + session.project.name)
saved_model = join(project_root, 'checkpoint/ddpg_' + session.project.name)
ddpg = DDPG(n_actions=knob_num, n_states=metric_num)
if os.path.exists(saved_memory):
ddpg.replay_memory.load_memory(saved_memory)
ddpg.load_model(saved_model)
ddpg = DDPG(n_actions=knob_num, n_states=metric_num, alr=ACTOR_LEARNING_RATE,
clr=CRITIC_LEARNING_RATE, gamma=GAMMA, batch_size=DDPG_BATCH_SIZE, tau=TAU)
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:
ddpg.replay_memory.set(session.ddpg_reply_memory)
knob_data = ddpg.choose_action(metric_data)
LOG.info('recommended knob: %s', knob_data)
knob_data = DataUtil.denormalize_knob_data(knob_data, knob_labels, session)
knob_bounds = np.vstack(DataUtil.get_knob_bounds(knob_labels, session))
knob_data = MinMaxScaler().fit(knob_bounds).inverse_transform(knob_data.reshape(1, -1))[0]
conf_map = {k: knob_data[i] for i, k in enumerate(knob_labels)}
conf_map_res = {}
conf_map_res['status'] = 'good'

41
server/website/website/utils.py
View File

@@ -93,30 +93,35 @@ class TaskUtil(object):
class DataUtil(object):
@staticmethod
def normalize_knob_data(knob_values, knob_labels, session):
for i, knob in enumerate(knob_labels):
def get_knob_bounds(knob_labels, session):
minvals = []
maxvals = []
for _, knob in enumerate(knob_labels):
knob_object = KnobCatalog.objects.get(dbms=session.dbms, name=knob, tunable=True)
minval = float(knob_object.minval)
maxval = float(knob_object.maxval)
knob_new = SessionKnob.objects.filter(knob=knob_object, session=session, tunable=True)
if knob_new.exists():
minval = float(knob_new[0].minval)
maxval = float(knob_new[0].maxval)
knob_values[i] = (knob_values[i] - minval) / (maxval - minval)
knob_values[i] = max(0, min(knob_values[i], 1))
return knob_values
@staticmethod
def denormalize_knob_data(knob_values, knob_labels, session):
for i, knob in enumerate(knob_labels):
knob_object = KnobCatalog.objects.get(dbms=session.dbms, name=knob, tunable=True)
minval = float(knob_object.minval)
maxval = float(knob_object.maxval)
knob_session_object = SessionKnob.objects.filter(knob=knob_object, session=session,
tunable=True)
if knob_session_object.exists():
minval = float(knob_session_object[0].minval)
maxval = float(knob_session_object[0].maxval)
else:
minval = float(knob_object.minval)
maxval = float(knob_object.maxval)
minvals.append(minval)
maxvals.append(maxval)
return np.array(minvals), np.array(maxvals)
@staticmethod
def denormalize_knob_data(knob_values, knob_labels, session):
for i, knob in enumerate(knob_labels):
knob_object = KnobCatalog.objects.get(dbms=session.dbms, name=knob, tunable=True)
knob_session_object = SessionKnob.objects.filter(knob=knob_object, session=session,
tunable=True)
if knob_session_object.exists():
minval = float(knob_session_object[0].minval)
maxval = float(knob_session_object[0].maxval)
else:
minval = float(knob_object.minval)
maxval = float(knob_object.maxval)
knob_values[i] = knob_values[i] * (maxval - minval) + minval
return knob_values

4
server/website/website/views.py
View File

@@ -30,8 +30,8 @@ from .models import (BackupData, DBMSCatalog, KnobCatalog, KnobData, MetricCatal
MetricData, MetricManager, Project, Result, Session, Workload,
SessionKnob)
from .parser import Parser
from .tasks import (aggregate_target_results, map_workload, train_ddpg, run_ddpg,
configuration_recommendation)
from .tasks import (aggregate_target_results, map_workload, train_ddpg,
configuration_recommendation, configuration_recommendation_ddpg)
from .types import (DBMSType, KnobUnitType, MetricType,
TaskType, VarType, WorkloadStatusType, AlgorithmType)
from .utils import JSONUtil, LabelUtil, MediaUtil, TaskUtil