ottertune/server/analysis/simulation.py

298 lines
10 KiB
Python
Raw Normal View History

#
# OtterTune - simulation.py
#
# Copyright (c) 2017-18, Carnegie Mellon University Database Group
#
2019-10-22 15:22:18 -07:00
2019-10-24 13:42:39 -07:00
import heapq
import random
import os
import sys
try:
import matplotlib.pyplot as plt
except (ModuleNotFoundError, ImportError):
plt = None
import numpy as np
import torch
sys.path.append("../")
from analysis.util import get_analysis_logger # noqa
from analysis.ddpg.ddpg import DDPG # noqa
2019-10-28 10:37:08 -07:00
from analysis.ddpg.ou_process import OUProcess # noqa
2019-10-24 13:42:39 -07:00
from analysis.gp_tf import GPRGD # noqa
from analysis.nn_tf import NeuralNet # noqa
LOG = get_analysis_logger(__name__)
class Environment(object):
2019-10-28 10:37:08 -07:00
def __init__(self, knob_dim, metric_dim, modes=[0], reward_variance=0,
metrics_variance=0.2):
self.knob_dim = knob_dim
self.metric_dim = metric_dim
self.modes = modes
self.mode = np.random.choice(self.modes)
self.counter = 0
self.reward_variance = reward_variance
self.metrics_variance = metrics_variance
def identity_sqrt(self, knob_data):
n1 = self.knob_dim // 4
n2 = self.knob_dim // 4
part1 = np.sum(knob_data[0: n1])
part2 = np.sum(np.sqrt(knob_data[n1: n1 + n2]))
reward = np.array([part1 + part2]) / (self.knob_dim // 2)
2019-10-28 10:37:08 -07:00
return reward
def threshold(self, knob_data):
n1 = self.knob_dim // 4
n2 = self.knob_dim // 4
part1 = np.sum(knob_data[0: n1] > 0.9)
part2 = np.sum(knob_data[n1: n1 + n2] < 0.1)
reward = np.array([part1 + part2]) / (self.knob_dim // 2)
return reward
def borehole(self, knob_data):
# ref: http://www.sfu.ca/~ssurjano/borehole.html
# pylint: disable=invalid-name
rw = knob_data[0] * (0.15 - 0.05) + 0.05
r = knob_data[1] * (50000 - 100) + 100
Tu = knob_data[2] * (115600 - 63070) + 63070
Hu = knob_data[3] * (1110 - 990) + 990
Tl = knob_data[4] * (116 - 63.1) + 63.1
Hl = knob_data[5] * (820 - 700) + 700
L = knob_data[6] * (1680 - 1120) + 1120
Kw = knob_data[7] * (12045 - 9855) + 9855
frac = 2 * L * Tu / (np.log(r / rw) * rw ** 2 * Kw)
2019-10-28 10:37:08 -07:00
reward = 2 * np.pi * Tu * (Hu - Hl) / (np.log(r / rw) * (1 + frac + Tu / Tl)) / 310
return np.array([reward])
def get_metrics(self, mode):
metrics = np.ones(self.metric_dim) * mode
metrics += np.random.rand(self.metric_dim) * self.metrics_variance
return metrics
def simulate_mode(self, knob_data, mode):
if mode == 0:
reward = self.identity_sqrt(knob_data)
elif mode == 1:
reward = self.threshold(knob_data)
elif mode == 2:
reward = np.zeros(1)
for i in range(0, len(knob_data), 8):
reward += self.borehole(knob_data[i: i+8])[0] / len(knob_data) * 8
reward = reward * (1.0 + self.reward_variance * np.random.rand(1)[0])
return reward, self.get_metrics(mode)
def simulate(self, knob_data):
2019-10-28 10:37:08 -07:00
self.counter += 1
k = 1
# every k runs, sample a new workload
if self.counter >= k:
self.counter = 0
self.mode = np.random.choice(self.modes)
return self.simulate_mode(knob_data, self.mode)
2019-10-28 10:37:08 -07:00
def ddpg(env, config, n_loops=100):
results = []
x_axis = []
2019-10-24 13:42:39 -07:00
gamma = config['gamma']
tau = config['tau']
2019-10-28 10:37:08 -07:00
a_lr = config['a_lr']
c_lr = config['c_lr']
2019-10-24 13:42:39 -07:00
n_epochs = config['n_epochs']
model_ddpg = DDPG(n_actions=env.knob_dim, n_states=env.metric_dim, gamma=gamma, tau=tau,
2019-10-28 10:37:08 -07:00
clr=c_lr, alr=a_lr)
knob_data = np.random.rand(env.knob_dim)
prev_metric_data = np.zeros(env.metric_dim)
2019-10-28 10:37:08 -07:00
for i in range(n_loops):
reward, metric_data = env.simulate(knob_data)
2019-10-28 10:37:08 -07:00
if i > 0:
model_ddpg.add_sample(prev_metric_data, prev_knob_data, prev_reward, metric_data)
prev_metric_data = metric_data
prev_knob_data = knob_data
prev_reward = reward
if i == 0:
continue
2019-10-24 13:42:39 -07:00
for _ in range(n_epochs):
model_ddpg.update()
results.append(reward)
2019-10-28 10:37:08 -07:00
x_axis.append(i+1)
LOG.info('loop: %d reward: %f', i, reward[0])
knob_data = model_ddpg.choose_action(metric_data)
return np.array(results), np.array(x_axis)
2019-10-24 13:42:39 -07:00
class ReplayMemory(object):
def __init__(self):
self.actions = []
self.rewards = []
def push(self, action, reward):
self.actions.append(action.tolist())
self.rewards.append(reward.tolist())
2019-10-24 13:42:39 -07:00
def get_all(self):
return self.actions, self.rewards
2019-10-24 13:42:39 -07:00
def dnn(env, config, n_loops=100):
results = []
x_axis = []
memory = ReplayMemory()
num_samples = config['num_samples']
2019-10-28 10:37:08 -07:00
ou_process = config['ou_process']
2019-10-24 13:42:39 -07:00
Xmin = np.zeros(env.knob_dim)
Xmax = np.ones(env.knob_dim)
2019-10-28 10:37:08 -07:00
noise = OUProcess(env.knob_dim)
2019-10-24 13:42:39 -07:00
for i in range(n_loops):
X_samples = np.random.rand(num_samples, env.knob_dim)
if i >= 10:
actions, rewards = memory.get_all()
tuples = tuple(zip(actions, rewards))
top10 = heapq.nlargest(10, tuples, key=lambda e: e[1])
for entry in top10:
X_samples = np.vstack((X_samples, np.array(entry[0])))
model_nn = NeuralNet(n_input=X_samples.shape[1],
batch_size=X_samples.shape[0],
2019-10-28 10:37:08 -07:00
learning_rate=0.01,
explore_iters=100,
2019-10-24 13:42:39 -07:00
noise_scale_begin=0.1,
noise_scale_end=0.0,
debug=False,
debug_interval=100)
if i >= 5:
actions, rewards = memory.get_all()
2019-10-28 10:37:08 -07:00
model_nn.fit(np.array(actions), -np.array(rewards), fit_epochs=50)
res = model_nn.recommend(X_samples, Xmin, Xmax, recommend_epochs=10, explore=False)
2019-10-24 13:42:39 -07:00
best_config_idx = np.argmin(res.minl.ravel())
best_config = res.minl_conf[best_config_idx, :]
2019-10-28 10:37:08 -07:00
if ou_process:
best_config += noise.noise()
best_config = best_config.clip(0, 1)
2019-10-24 13:42:39 -07:00
reward, _ = env.simulate(best_config)
memory.push(best_config, reward)
LOG.info('loop: %d reward: %f', i, reward[0])
results.append(reward)
2019-10-28 10:37:08 -07:00
x_axis.append(i+1)
2019-10-24 13:42:39 -07:00
return np.array(results), np.array(x_axis)
def gprgd(env, config, n_loops=100):
results = []
x_axis = []
memory = ReplayMemory()
2019-10-28 10:37:08 -07:00
num_collections = config['num_collections']
2019-10-24 13:42:39 -07:00
num_samples = config['num_samples']
X_min = np.zeros(env.knob_dim)
X_max = np.ones(env.knob_dim)
2019-10-28 10:37:08 -07:00
for _ in range(num_collections):
2019-10-24 13:42:39 -07:00
action = np.random.rand(env.knob_dim)
reward, _ = env.simulate(action)
memory.push(action, reward)
for i in range(n_loops):
X_samples = np.random.rand(num_samples, env.knob_dim)
2019-10-28 10:37:08 -07:00
if i >= 10:
2019-10-24 13:42:39 -07:00
actions, rewards = memory.get_all()
tuples = tuple(zip(actions, rewards))
top10 = heapq.nlargest(10, tuples, key=lambda e: e[1])
for entry in top10:
# Tensorflow get broken if we use the training data points as
# starting points for GPRGD.
X_samples = np.vstack((X_samples, np.array(entry[0]) * 0.97 + 0.01))
model = GPRGD(length_scale=1.0,
magnitude=1.0,
2019-10-28 10:37:08 -07:00
max_train_size=100,
batch_size=100,
2019-10-24 13:42:39 -07:00
num_threads=4,
learning_rate=0.01,
epsilon=1e-6,
max_iter=500,
2019-10-28 10:37:08 -07:00
sigma_multiplier=30.0,
2019-10-24 13:42:39 -07:00
mu_multiplier=1.0)
actions, rewards = memory.get_all()
model.fit(np.array(actions), -np.array(rewards), X_min, X_max, ridge=0.01)
res = model.predict(X_samples)
best_config_idx = np.argmin(res.minl.ravel())
best_config = res.minl_conf[best_config_idx, :]
reward, _ = env.simulate(best_config)
memory.push(best_config, reward)
LOG.info('loop: %d reward: %f', i, reward[0])
results.append(reward)
x_axis.append(i)
return np.array(results), np.array(x_axis)
2019-10-28 10:37:08 -07:00
def plotlines(xs, results, labels, title, path):
if plt:
2019-10-28 10:37:08 -07:00
figsize = 13, 10
figure, ax = plt.subplots(figsize=figsize)
lines = []
N = 20
weights = np.ones(N)
for x_axis, result, label in zip(xs, results, labels):
result = np.convolve(weights/weights.sum(), result.flatten())[N-1:-N+1]
lines.append(plt.plot(x_axis[:-N+1], result, label=label, lw=4)[0])
plt.legend(handles=lines, fontsize=30)
plt.title(title, fontsize=25)
plt.xticks(fontsize=25)
plt.yticks(fontsize=25)
ax.set_xlabel("loops", fontsize=30)
ax.set_ylabel("rewards", fontsize=30)
plt.savefig(path)
plt.clf()
2019-10-28 10:37:08 -07:00
def run(tuners, configs, labels, title, env, n_loops, n_repeats):
if not plt:
LOG.info("Cannot import matplotlib. Will write results to files instead of figures.")
random.seed(0)
2019-10-28 10:37:08 -07:00
np.random.seed(1)
torch.manual_seed(0)
2019-10-24 13:42:39 -07:00
results = []
2019-10-28 10:37:08 -07:00
xs = []
for j, _ in enumerate(tuners):
for i in range(n_repeats[j]):
2019-10-24 13:42:39 -07:00
result, x_axis = tuners[j](env, configs[j], n_loops=n_loops)
if i is 0:
2019-10-28 10:37:08 -07:00
results.append(result / n_repeats[j])
xs.append(x_axis)
2019-10-24 13:42:39 -07:00
else:
2019-10-28 10:37:08 -07:00
results[j] += result / n_repeats[j]
2019-10-24 13:42:39 -07:00
if plt:
2019-10-28 10:37:08 -07:00
if not os.path.exists("simulation_figures"):
os.mkdir("simulation_figures")
filename = "simulation_figures/{}.pdf".format(title)
plotlines(xs, results, labels, title, filename)
if not os.path.exists("simulation_results"):
os.mkdir("simulation_results")
2019-10-24 13:42:39 -07:00
for j in range(len(tuners)):
2019-10-28 10:37:08 -07:00
with open("simulation_results/" + title + '_' + labels[j] + '.csv', 'w') as f:
for i, result in zip(xs[j], results[j]):
2019-10-24 13:42:39 -07:00
f.write(str(i) + ',' + str(result[0]) + '\n')
def main():
2019-10-28 10:37:08 -07:00
env = Environment(knob_dim=192, metric_dim=60, modes=[0, 1], reward_variance=0.05)
n_loops = 2000
configs = [{'gamma': 0, 'tau': 0.002, 'a_lr': 0.01, 'c_lr': 0.01, 'n_epochs': 1},
{'gamma': 0, 'tau': 0.002, 'a_lr': 0.01, 'c_lr': 0.001, 'n_epochs': 1},
{'gamma': 0., 'tau': 0.002, 'a_lr': 0.001, 'c_lr': 0.001, 'n_epochs': 1},
# {'num_samples': 100, 'ou_process': False},
]
tuners = [ddpg, ddpg, ddpg]
labels = ['1', '2', '3']
title = 'varing_workloads'
n_repeats = [3, 3, 3]
run(tuners, configs, labels, title, env, n_loops, n_repeats)
if __name__ == '__main__':
main()