def test_trace_recency(self):
# Ensure that config for unseen recency is passed to unseen items, first observation will always be unseen
env = park.make('cache')
config.cache_unseen_recency = 1000
obs = env.reset()
self.assertTrue(obs[2] == 1000)
config.cache_unseen_recency = 500
obs = env.reset()
self.assertTrue(obs[2] == 500)
def run_env_with_random_agent(env_name, seed):
# suppress unittest from throwing weird warnings
warnings.simplefilter('ignore', category=ImportWarning)
env = park.make(env_name)
env.seed(seed)
obs = env.reset()
done = False
while not done:
act = env.action_space.sample()
obs, reward, done, info = env.step(act)
def _thunk():
if env_id not in PARK_ENV_LIST:
raise ValueError(
"Unsupported environment, expect the environment to be one of "
+ str(PARK_ENV_LIST) + " but got: " + str(env_id))
elif env_id == 'load_balance':
# arrange the number of stream jobs
env = park.make(env_id,
num_stream_jobs=args.num_stream_jobs,
service_rates=args.load_balance_service_rates)
# random act after resetting to diversify the state
# only use when training
if train:
env = LoadBalanceRandomReset(env, args.max_random_init_steps)
# if using load balance, clip and normalize the observation with this wrapper
if args is not None:
env = ProcessLoadBalanceObservation(
env, args.job_size_norm_factor,
args.server_load_norm_factor, args.highest_server_obs,
args.highest_job_obs, args.elapsed_time_norm_factor,
args.highest_elapsed_time)
# normalize reward
env = RewardNormalize(env, args.reward_norm_factor)
if args.fix_job_sequence:
# fix job sequence
env = FixJobSequence(env, seed)
if max_episode_steps:
env = TimeLimit(env, max_episode_steps)
# adding information to env for computing return
env = TimeLimitMask(env)
# IMPORTANT: all environments used same random seed to repeat the input-process
if train and args.algo.startswith('mib'):
env.seed(seed)
else:
env.seed(seed + rank)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
return env
def __init__(self):
"""
working now:
1. load_balance
2. abr_sim
3. switch_scheduling
4. simple_queue
5. region_assignment
not working:
1. abr
2. aqm
3. congestion_control
4. spark_sim
5. query_optimizer
6. cache
7. tf_placement
8. circuit_three_stage_transimpedance
9. tf_placement_sim
10. multi_dim_index
11. spark
"""
self.env = park.make('abr_sim')
def train(self,
db,
training_samples,
use_subqueries=False,
test_samples=None):
assert isinstance(training_samples[0], dict)
if not self.nn_type == "num_tables":
self.num_threads = multiprocessing.cpu_count()
print("setting num threads to: ", self.num_threads)
torch.set_num_threads(self.num_threads)
else:
self.num_threads = -1
self.db = db
db.init_featurizer(num_tables_feature=self.num_tables_feature,
max_discrete_featurizing_buckets=self.
max_discrete_featurizing_buckets)
# get one true source of X,Y feature vector pairs, which won't be
# reused.
start = time.time()
self.training_samples = training_samples
self.Xtrain, self.Ytrain, self.train_num_table_mapping = \
self._get_feature_vectors(self.training_samples)
# create a new park env, and close at the end.
self.env = park.make('query_optimizer')
self.test_samples = test_samples
if test_samples is not None and len(test_samples) > 0:
# random.shuffle(test_samples)
self.Xtest, self.Ytest, self.test_num_table_mapping = \
self._get_feature_vectors(self.test_samples)
print("{} training, {} test subqueries".format(
len(self.Xtrain), len(self.Xtest)))
self.test_env = park.make('query_optimizer')
print("feature len: {}, generation time: {}".\
format(len(self.Xtrain[0]), time.time()-start))
# FIXME: multiple table version
self.init_nets()
model_size = self.num_parameters()
print("model size: {} MB".format(model_size))
self.num_iter = 0
# start off uniformly
self.subquery_sampling_weights = [1 / len(self.Xtrain)] * len(
self.Xtrain)
prev_end = time.time()
while True:
if (self.num_iter % 100 == 0):
pass
# progress stuff
it_time = time.time() - prev_end
prev_end = time.time()
print("MB: {}, T:{}, I:{} : {}".format(\
self.mb_size, self.num_threads, self.num_iter, it_time))
sys.stdout.flush()
if (self.num_iter % self.eval_iter == 0):
# we will wait on these results when we reach this point in the
# next iteration
self._update_join_results(self.train_join_results,
self.training_samples, "train",
self.num_iter - self.eval_iter)
self.train_join_results = self._periodic_eval(
self.Xtrain, self.Ytrain, self.training_samples, "train",
self.train_join_loss_pool, self.env)
# TODO: handle reweighing schemes here
if test_samples is not None:
self._update_join_results(self.test_join_results,
self.test_samples, "test",
self.num_iter - self.eval_iter)
self.test_join_results = self._periodic_eval(
self.Xtest, self.Ytest, self.test_samples, "test",
self.test_join_loss_pool, self.test_env)
if self.num_iter % self.eval_iter == 0:
self.save_stats()
self.train_step(self.eval_iter)
self.num_iter += self.eval_iter
if (self.num_iter >= self.max_iter):
print("breaking because max iter done")
break
for steps in range(num_steps):
action, _ = agent.predict(state)
new_state, reward, done, _ = env.step(action)
rewards.append(reward)
state = new_state
if done or steps == num_steps - 1:
all_rewards.append(np.sum(rewards))
all_lengths.append(steps)
average_lengths.append(np.mean(all_lengths[-10:]))
if episode % 1 == 0:
print(
"episode: {}, reward: {}, total length: {}, average length: {} \n"
.format(episode, np.sum(rewards), steps,
average_lengths[-1]))
print(
"count_bhr: {}, count_ohr: {}, total_size: {}, bhr_ratio: {} \n"
.format(
env.sim.count_bhr, env.sim.count_ohr,
env.sim.size_all,
float(
float(env.sim.count_bhr) /
float(env.sim.size_all))))
break
env = park.make('cache')
admitAll(env)
def test_bounds_low(self):
# New lower bound for the cache test traces, test trace numbers start at 0
env = park.make('cache')
env.reset(low=0, high=1)
def test_bounds(self):
# New upper bound for the cache test traces, test trace numbers end at 999
env = park.make('cache')
env.reset(low=1000, high=1001)
def main():
env = park.make('congestion_control')
env.run(RandomAgent, ())
