def test_inverse_transform(self):
X = [[1, 0, 2], [1, 1, 2], [1, 2, 2]]
n_values = [3]
categorical_features = [1]
cat_columnlabels = ['label']
noncat_columnlabels = ['a', 'b']
X_expected = [[1, 0, 0, 1, 2], [0, 1, 0, 1, 2], [0, 0, 1, 1, 2]]
enc = DummyEncoder(n_values, categorical_features, cat_columnlabels,
noncat_columnlabels)
X_encoded = enc.fit_transform(X)
self.assertTrue(np.all(X_encoded == X_expected))
X_decoded = enc.inverse_transform(X_encoded)
self.assertTrue(np.all(X == X_decoded))
def configuration_recommendation(target_data):
LOG.info('configuration_recommendation called')
latest_pipeline_run = PipelineRun.objects.get_latest()
if target_data['bad'] is True:
target_data_res = {}
target_data_res['status'] = 'bad'
target_data_res[
'info'] = 'WARNING: no training data, the config is generated randomly'
target_data_res['recommendation'] = target_data['config_recommend']
return target_data_res
# Load mapped workload data
mapped_workload_id = target_data['mapped_workload'][0]
mapped_workload = Workload.objects.get(pk=mapped_workload_id)
workload_knob_data = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.KNOB_DATA)
workload_knob_data = JSONUtil.loads(workload_knob_data.data)
workload_metric_data = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.METRIC_DATA)
workload_metric_data = JSONUtil.loads(workload_metric_data.data)
X_workload = np.array(workload_knob_data['data'])
X_columnlabels = np.array(workload_knob_data['columnlabels'])
y_workload = np.array(workload_metric_data['data'])
y_columnlabels = np.array(workload_metric_data['columnlabels'])
rowlabels_workload = np.array(workload_metric_data['rowlabels'])
# Target workload data
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
X_target = target_data['X_matrix']
y_target = target_data['y_matrix']
rowlabels_target = np.array(target_data['rowlabels'])
if not np.array_equal(X_columnlabels, target_data['X_columnlabels']):
raise Exception(('The workload and target data should have '
'identical X columnlabels (sorted knob names)'))
if not np.array_equal(y_columnlabels, target_data['y_columnlabels']):
raise Exception(('The workload and target data should have '
'identical y columnlabels (sorted metric names)'))
# Filter Xs by top 10 ranked knobs
ranked_knobs = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.RANKED_KNOBS)
ranked_knobs = JSONUtil.loads(ranked_knobs.data)[:IMPORTANT_KNOB_NUMBER]
ranked_knob_idxs = [
i for i, cl in enumerate(X_columnlabels) if cl in ranked_knobs
]
X_workload = X_workload[:, ranked_knob_idxs]
X_target = X_target[:, ranked_knob_idxs]
X_columnlabels = X_columnlabels[ranked_knob_idxs]
# Filter ys by current target objective metric
target_objective = newest_result.session.target_objective
target_obj_idx = [
i for i, cl in enumerate(y_columnlabels) if cl == target_objective
]
if len(target_obj_idx) == 0:
raise Exception(('Could not find target objective in metrics '
'(target_obj={})').format(target_objective))
elif len(target_obj_idx) > 1:
raise Exception(
('Found {} instances of target objective in '
'metrics (target_obj={})').format(len(target_obj_idx),
target_objective))
metric_meta = MetricCatalog.objects.get_metric_meta(
newest_result.session.dbms, newest_result.session.target_objective)
if metric_meta[target_objective].improvement == '(less is better)':
lessisbetter = True
else:
lessisbetter = False
y_workload = y_workload[:, target_obj_idx]
y_target = y_target[:, target_obj_idx]
y_columnlabels = y_columnlabels[target_obj_idx]
# Combine duplicate rows in the target/workload data (separately)
X_workload, y_workload, rowlabels_workload = DataUtil.combine_duplicate_rows(
X_workload, y_workload, rowlabels_workload)
X_target, y_target, rowlabels_target = DataUtil.combine_duplicate_rows(
X_target, y_target, rowlabels_target)
# Delete any rows that appear in both the workload data and the target
# data from the workload data
dups_filter = np.ones(X_workload.shape[0], dtype=bool)
target_row_tups = [tuple(row) for row in X_target]
for i, row in enumerate(X_workload):
if tuple(row) in target_row_tups:
dups_filter[i] = False
X_workload = X_workload[dups_filter, :]
y_workload = y_workload[dups_filter, :]
rowlabels_workload = rowlabels_workload[dups_filter]
# Combine target & workload Xs for preprocessing
X_matrix = np.vstack([X_target, X_workload])
# Dummy encode categorial variables
categorical_info = DataUtil.dummy_encoder_helper(X_columnlabels,
mapped_workload.dbms)
dummy_encoder = DummyEncoder(categorical_info['n_values'],
categorical_info['categorical_features'],
categorical_info['cat_columnlabels'],
categorical_info['noncat_columnlabels'])
X_matrix = dummy_encoder.fit_transform(X_matrix)
# below two variables are needed for correctly determing max/min on dummies
binary_index_set = set(categorical_info['binary_vars'])
total_dummies = dummy_encoder.total_dummies()
# Scale to N(0, 1)
X_scaler = StandardScaler()
X_scaled = X_scaler.fit_transform(X_matrix)
if y_target.shape[0] < 5: # FIXME
# FIXME (dva): if there are fewer than 5 target results so far
# then scale the y values (metrics) using the workload's
# y_scaler. I'm not sure if 5 is the right cutoff.
y_target_scaler = None
y_workload_scaler = StandardScaler()
y_matrix = np.vstack([y_target, y_workload])
y_scaled = y_workload_scaler.fit_transform(y_matrix)
else:
# FIXME (dva): otherwise try to compute a separate y_scaler for
# the target and scale them separately.
try:
y_target_scaler = StandardScaler()
y_workload_scaler = StandardScaler()
y_target_scaled = y_target_scaler.fit_transform(y_target)
y_workload_scaled = y_workload_scaler.fit_transform(y_workload)
y_scaled = np.vstack([y_target_scaled, y_workload_scaled])
except ValueError:
y_target_scaler = None
y_workload_scaler = StandardScaler()
y_scaled = y_workload_scaler.fit_transform(y_target)
# Set up constraint helper
constraint_helper = ParamConstraintHelper(
scaler=X_scaler,
encoder=dummy_encoder,
binary_vars=categorical_info['binary_vars'],
init_flip_prob=INIT_FLIP_PROB,
flip_prob_decay=FLIP_PROB_DECAY)
# FIXME (dva): check if these are good values for the ridge
# ridge = np.empty(X_scaled.shape[0])
# ridge[:X_target.shape[0]] = 0.01
# ridge[X_target.shape[0]:] = 0.1
# FIXME: we should generate more samples and use a smarter sampling
# technique
num_samples = NUM_SAMPLES
X_samples = np.empty((num_samples, X_scaled.shape[1]))
X_min = np.empty(X_scaled.shape[1])
X_max = np.empty(X_scaled.shape[1])
knobs_mem = KnobCatalog.objects.filter(dbms=newest_result.session.dbms,
tunable=True,
resource=1)
knobs_mem_catalog = {k.name: k for k in knobs_mem}
mem_max = newest_result.workload.hardware.memory
X_mem = np.zeros([1, X_scaled.shape[1]])
X_default = np.empty(X_scaled.shape[1])
# Get default knob values
for i, k_name in enumerate(X_columnlabels):
k = KnobCatalog.objects.filter(dbms=newest_result.session.dbms,
name=k_name)[0]
X_default[i] = k.default
X_default_scaled = X_scaler.transform(
X_default.reshape(1, X_default.shape[0]))[0]
# Determine min/max for knob values
for i in range(X_scaled.shape[1]):
if i < total_dummies or i in binary_index_set:
col_min = 0
col_max = 1
else:
col_min = X_scaled[:, i].min()
col_max = X_scaled[:, i].max()
if X_columnlabels[i] in knobs_mem_catalog:
X_mem[0][i] = mem_max * 1024 * 1024 * 1024 # mem_max GB
col_max = min(col_max, X_scaler.transform(X_mem)[0][i])
# Set min value to the default value
# FIXME: support multiple methods can be selected by users
col_min = X_default_scaled[i]
X_min[i] = col_min
X_max[i] = col_max
X_samples[:, i] = np.random.rand(num_samples) * (col_max -
col_min) + col_min
# Maximize the throughput, moreisbetter
# Use gradient descent to minimize -throughput
if not lessisbetter:
y_scaled = -y_scaled
q = queue.PriorityQueue()
for x in range(0, y_scaled.shape[0]):
q.put((y_scaled[x][0], x))
i = 0
while i < TOP_NUM_CONFIG:
try:
item = q.get_nowait()
# Tensorflow get broken if we use the training data points as
# starting points for GPRGD. We add a small bias for the
# starting points. GPR_EPS default value is 0.001
# if the starting point is X_max, we minus a small bias to
# make sure it is within the range.
dist = sum(np.square(X_max - X_scaled[item[1]]))
if dist < 0.001:
X_samples = np.vstack(
(X_samples, X_scaled[item[1]] - abs(GPR_EPS)))
else:
X_samples = np.vstack(
(X_samples, X_scaled[item[1]] + abs(GPR_EPS)))
i = i + 1
except queue.Empty:
break
model = GPRGD(length_scale=DEFAULT_LENGTH_SCALE,
magnitude=DEFAULT_MAGNITUDE,
max_train_size=MAX_TRAIN_SIZE,
batch_size=BATCH_SIZE,
num_threads=NUM_THREADS,
learning_rate=DEFAULT_LEARNING_RATE,
epsilon=DEFAULT_EPSILON,
max_iter=MAX_ITER,
sigma_multiplier=DEFAULT_SIGMA_MULTIPLIER,
mu_multiplier=DEFAULT_MU_MULTIPLIER)
model.fit(X_scaled, y_scaled, X_min, X_max, ridge=DEFAULT_RIDGE)
res = model.predict(X_samples, constraint_helper=constraint_helper)
best_config_idx = np.argmin(res.minl.ravel())
best_config = res.minl_conf[best_config_idx, :]
best_config = X_scaler.inverse_transform(best_config)
# Decode one-hot encoding into categorical knobs
best_config = dummy_encoder.inverse_transform(best_config)
# Although we have max/min limits in the GPRGD training session, it may
# lose some precisions. e.g. 0.99..99 >= 1.0 may be True on the scaled data,
# when we inversely transform the scaled data, the different becomes much larger
# and cannot be ignored. Here we check the range on the original data
# directly, and make sure the recommended config lies within the range
X_min_inv = X_scaler.inverse_transform(X_min)
X_max_inv = X_scaler.inverse_transform(X_max)
best_config = np.minimum(best_config, X_max_inv)
best_config = np.maximum(best_config, X_min_inv)
conf_map = {k: best_config[i] for i, k in enumerate(X_columnlabels)}
conf_map_res = {}
conf_map_res['status'] = 'good'
conf_map_res['recommendation'] = conf_map
conf_map_res['info'] = 'INFO: training data size is {}'.format(
X_scaled.shape[0])
return conf_map_res
def configuration_recommendation(recommendation_input):
target_data, algorithm = recommendation_input
LOG.info('configuration_recommendation called')
if target_data['bad'] is True:
target_data_res = dict(
status='bad',
result_id=target_data['newest_result_id'],
info='WARNING: no training data, the config is generated randomly',
recommendation=target_data['config_recommend'],
pipeline_run=target_data['pipeline_run'])
LOG.debug('%s: Skipping configuration recommendation.\n\ndata=%s\n',
AlgorithmType.name(algorithm),
JSONUtil.dumps(target_data, pprint=True))
return target_data_res
# Load mapped workload data
mapped_workload_id = target_data['mapped_workload'][0]
latest_pipeline_run = PipelineRun.objects.get(
pk=target_data['pipeline_run'])
mapped_workload = Workload.objects.get(pk=mapped_workload_id)
workload_knob_data = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.KNOB_DATA)
workload_knob_data = JSONUtil.loads(workload_knob_data.data)
workload_metric_data = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.METRIC_DATA)
workload_metric_data = JSONUtil.loads(workload_metric_data.data)
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
cleaned_workload_knob_data = clean_knob_data(
workload_knob_data["data"], workload_knob_data["columnlabels"],
newest_result.session)
X_workload = np.array(cleaned_workload_knob_data[0])
X_columnlabels = np.array(cleaned_workload_knob_data[1])
y_workload = np.array(workload_metric_data['data'])
y_columnlabels = np.array(workload_metric_data['columnlabels'])
rowlabels_workload = np.array(workload_metric_data['rowlabels'])
# Target workload data
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
X_target = target_data['X_matrix']
y_target = target_data['y_matrix']
rowlabels_target = np.array(target_data['rowlabels'])
if not np.array_equal(X_columnlabels, target_data['X_columnlabels']):
raise Exception(('The workload and target data should have '
'identical X columnlabels (sorted knob names)'))
if not np.array_equal(y_columnlabels, target_data['y_columnlabels']):
raise Exception(('The workload and target data should have '
'identical y columnlabels (sorted metric names)'))
# Filter Xs by top 10 ranked knobs
ranked_knobs = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.RANKED_KNOBS)
ranked_knobs = JSONUtil.loads(ranked_knobs.data)[:IMPORTANT_KNOB_NUMBER]
ranked_knob_idxs = [
i for i, cl in enumerate(X_columnlabels) if cl in ranked_knobs
]
X_workload = X_workload[:, ranked_knob_idxs]
X_target = X_target[:, ranked_knob_idxs]
X_columnlabels = X_columnlabels[ranked_knob_idxs]
# Filter ys by current target objective metric
target_objective = newest_result.session.target_objective
target_obj_idx = [
i for i, cl in enumerate(y_columnlabels) if cl == target_objective
]
if len(target_obj_idx) == 0:
raise Exception(('Could not find target objective in metrics '
'(target_obj={})').format(target_objective))
elif len(target_obj_idx) > 1:
raise Exception(
('Found {} instances of target objective in '
'metrics (target_obj={})').format(len(target_obj_idx),
target_objective))
metric_meta = db.target_objectives.get_metric_metadata(
newest_result.session.dbms.pk, newest_result.session.target_objective)
lessisbetter = metric_meta[
target_objective].improvement == db.target_objectives.LESS_IS_BETTER
y_workload = y_workload[:, target_obj_idx]
y_target = y_target[:, target_obj_idx]
y_columnlabels = y_columnlabels[target_obj_idx]
# Combine duplicate rows in the target/workload data (separately)
X_workload, y_workload, rowlabels_workload = DataUtil.combine_duplicate_rows(
X_workload, y_workload, rowlabels_workload)
X_target, y_target, rowlabels_target = DataUtil.combine_duplicate_rows(
X_target, y_target, rowlabels_target)
# Delete any rows that appear in both the workload data and the target
# data from the workload data
dups_filter = np.ones(X_workload.shape[0], dtype=bool)
target_row_tups = [tuple(row) for row in X_target]
for i, row in enumerate(X_workload):
if tuple(row) in target_row_tups:
dups_filter[i] = False
X_workload = X_workload[dups_filter, :]
y_workload = y_workload[dups_filter, :]
rowlabels_workload = rowlabels_workload[dups_filter]
# Combine target & workload Xs for preprocessing
X_matrix = np.vstack([X_target, X_workload])
# Dummy encode categorial variables
categorical_info = DataUtil.dummy_encoder_helper(X_columnlabels,
mapped_workload.dbms)
dummy_encoder = DummyEncoder(categorical_info['n_values'],
categorical_info['categorical_features'],
categorical_info['cat_columnlabels'],
categorical_info['noncat_columnlabels'])
X_matrix = dummy_encoder.fit_transform(X_matrix)
# below two variables are needed for correctly determing max/min on dummies
binary_index_set = set(categorical_info['binary_vars'])
total_dummies = dummy_encoder.total_dummies()
# Scale to N(0, 1)
X_scaler = StandardScaler()
X_scaled = X_scaler.fit_transform(X_matrix)
if y_target.shape[0] < 5: # FIXME
# FIXME (dva): if there are fewer than 5 target results so far
# then scale the y values (metrics) using the workload's
# y_scaler. I'm not sure if 5 is the right cutoff.
y_target_scaler = None
y_workload_scaler = StandardScaler()
y_matrix = np.vstack([y_target, y_workload])
y_scaled = y_workload_scaler.fit_transform(y_matrix)
else:
# FIXME (dva): otherwise try to compute a separate y_scaler for
# the target and scale them separately.
try:
y_target_scaler = StandardScaler()
y_workload_scaler = StandardScaler()
y_target_scaled = y_target_scaler.fit_transform(y_target)
y_workload_scaled = y_workload_scaler.fit_transform(y_workload)
y_scaled = np.vstack([y_target_scaled, y_workload_scaled])
except ValueError:
y_target_scaler = None
y_workload_scaler = StandardScaler()
y_scaled = y_workload_scaler.fit_transform(y_target)
# Set up constraint helper
constraint_helper = ParamConstraintHelper(
scaler=X_scaler,
encoder=dummy_encoder,
binary_vars=categorical_info['binary_vars'],
init_flip_prob=INIT_FLIP_PROB,
flip_prob_decay=FLIP_PROB_DECAY)
# FIXME (dva): check if these are good values for the ridge
# ridge = np.empty(X_scaled.shape[0])
# ridge[:X_target.shape[0]] = 0.01
# ridge[X_target.shape[0]:] = 0.1
# FIXME: we should generate more samples and use a smarter sampling
# technique
num_samples = NUM_SAMPLES
X_samples = np.empty((num_samples, X_scaled.shape[1]))
X_min = np.empty(X_scaled.shape[1])
X_max = np.empty(X_scaled.shape[1])
X_scaler_matrix = np.zeros([1, X_scaled.shape[1]])
session_knobs = SessionKnob.objects.get_knobs_for_session(
newest_result.session)
# Set min/max for knob values
for i in range(X_scaled.shape[1]):
if i < total_dummies or i in binary_index_set:
col_min = 0
col_max = 1
else:
col_min = X_scaled[:, i].min()
col_max = X_scaled[:, i].max()
for knob in session_knobs:
if X_columnlabels[i] == knob["name"]:
X_scaler_matrix[0][i] = knob["minval"]
col_min = X_scaler.transform(X_scaler_matrix)[0][i]
X_scaler_matrix[0][i] = knob["maxval"]
col_max = X_scaler.transform(X_scaler_matrix)[0][i]
X_min[i] = col_min
X_max[i] = col_max
X_samples[:, i] = np.random.rand(num_samples) * (col_max -
col_min) + col_min
# Maximize the throughput, moreisbetter
# Use gradient descent to minimize -throughput
if not lessisbetter:
y_scaled = -y_scaled
q = queue.PriorityQueue()
for x in range(0, y_scaled.shape[0]):
q.put((y_scaled[x][0], x))
i = 0
while i < TOP_NUM_CONFIG:
try:
item = q.get_nowait()
# Tensorflow get broken if we use the training data points as
# starting points for GPRGD. We add a small bias for the
# starting points. GPR_EPS default value is 0.001
# if the starting point is X_max, we minus a small bias to
# make sure it is within the range.
dist = sum(np.square(X_max - X_scaled[item[1]]))
if dist < 0.001:
X_samples = np.vstack(
(X_samples, X_scaled[item[1]] - abs(GPR_EPS)))
else:
X_samples = np.vstack(
(X_samples, X_scaled[item[1]] + abs(GPR_EPS)))
i = i + 1
except queue.Empty:
break
session = newest_result.session
res = None
if algorithm == AlgorithmType.DNN:
# neural network model
model_nn = NeuralNet(n_input=X_samples.shape[1],
batch_size=X_samples.shape[0],
explore_iters=DNN_EXPLORE_ITER,
noise_scale_begin=DNN_NOISE_SCALE_BEGIN,
noise_scale_end=DNN_NOISE_SCALE_END,
debug=DNN_DEBUG,
debug_interval=DNN_DEBUG_INTERVAL)
if session.dnn_model is not None:
model_nn.set_weights_bin(session.dnn_model)
model_nn.fit(X_scaled, y_scaled, fit_epochs=DNN_TRAIN_ITER)
res = model_nn.recommend(X_samples,
X_min,
X_max,
explore=DNN_EXPLORE,
recommend_epochs=MAX_ITER)
session.dnn_model = model_nn.get_weights_bin()
session.save()
elif algorithm == AlgorithmType.GPR:
# default gpr model
model = GPRGD(length_scale=DEFAULT_LENGTH_SCALE,
magnitude=DEFAULT_MAGNITUDE,
max_train_size=MAX_TRAIN_SIZE,
batch_size=BATCH_SIZE,
num_threads=NUM_THREADS,
learning_rate=DEFAULT_LEARNING_RATE,
epsilon=DEFAULT_EPSILON,
max_iter=MAX_ITER,
sigma_multiplier=DEFAULT_SIGMA_MULTIPLIER,
mu_multiplier=DEFAULT_MU_MULTIPLIER)
model.fit(X_scaled, y_scaled, X_min, X_max, ridge=DEFAULT_RIDGE)
res = model.predict(X_samples, constraint_helper=constraint_helper)
best_config_idx = np.argmin(res.minl.ravel())
best_config = res.minl_conf[best_config_idx, :]
best_config = X_scaler.inverse_transform(best_config)
# Decode one-hot encoding into categorical knobs
best_config = dummy_encoder.inverse_transform(best_config)
# Although we have max/min limits in the GPRGD training session, it may
# lose some precisions. e.g. 0.99..99 >= 1.0 may be True on the scaled data,
# when we inversely transform the scaled data, the different becomes much larger
# and cannot be ignored. Here we check the range on the original data
# directly, and make sure the recommended config lies within the range
X_min_inv = X_scaler.inverse_transform(X_min)
X_max_inv = X_scaler.inverse_transform(X_max)
best_config = np.minimum(best_config, X_max_inv)
best_config = np.maximum(best_config, X_min_inv)
conf_map = {k: best_config[i] for i, k in enumerate(X_columnlabels)}
conf_map_res = dict(status='good',
result_id=target_data['newest_result_id'],
recommendation=conf_map,
info='INFO: training data size is {}'.format(
X_scaled.shape[0]),
pipeline_run=latest_pipeline_run.pk)
LOG.debug('%s: Finished selecting the next config.\n\ndata=%s\n',
AlgorithmType.name(algorithm),
JSONUtil.dumps(conf_map_res, pprint=True))
return conf_map_res
def configuration_recommendation(target_data):
LOG.info('configuration_recommendation called')
latest_pipeline_run = PipelineRun.objects.get_latest()
if target_data['bad'] is True:
target_data_res = {}
target_data_res['status'] = 'bad'
target_data_res['info'] = 'WARNING: no training data, the config is generated randomly'
target_data_res['recommendation'] = target_data['config_recommend']
return target_data_res
# Load mapped workload data
mapped_workload_id = target_data['mapped_workload'][0]
mapped_workload = Workload.objects.get(pk=mapped_workload_id)
workload_knob_data = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.KNOB_DATA)
workload_knob_data = JSONUtil.loads(workload_knob_data.data)
workload_metric_data = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.METRIC_DATA)
workload_metric_data = JSONUtil.loads(workload_metric_data.data)
X_workload = np.array(workload_knob_data['data'])
X_columnlabels = np.array(workload_knob_data['columnlabels'])
y_workload = np.array(workload_metric_data['data'])
y_columnlabels = np.array(workload_metric_data['columnlabels'])
rowlabels_workload = np.array(workload_metric_data['rowlabels'])
# Target workload data
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
X_target = target_data['X_matrix']
y_target = target_data['y_matrix']
rowlabels_target = np.array(target_data['rowlabels'])
if not np.array_equal(X_columnlabels, target_data['X_columnlabels']):
raise Exception(('The workload and target data should have '
'identical X columnlabels (sorted knob names)'))
if not np.array_equal(y_columnlabels, target_data['y_columnlabels']):
raise Exception(('The workload and target data should have '
'identical y columnlabels (sorted metric names)'))
# Filter Xs by top 10 ranked knobs
ranked_knobs = PipelineData.objects.get(
pipeline_run=latest_pipeline_run,
workload=mapped_workload,
task_type=PipelineTaskType.RANKED_KNOBS)
ranked_knobs = JSONUtil.loads(ranked_knobs.data)[:IMPORTANT_KNOB_NUMBER]
ranked_knob_idxs = [i for i, cl in enumerate(X_columnlabels) if cl in ranked_knobs]
X_workload = X_workload[:, ranked_knob_idxs]
X_target = X_target[:, ranked_knob_idxs]
X_columnlabels = X_columnlabels[ranked_knob_idxs]
# Filter ys by current target objective metric
target_objective = newest_result.session.target_objective
target_obj_idx = [i for i, cl in enumerate(y_columnlabels) if cl == target_objective]
if len(target_obj_idx) == 0:
raise Exception(('Could not find target objective in metrics '
'(target_obj={})').format(target_objective))
elif len(target_obj_idx) > 1:
raise Exception(('Found {} instances of target objective in '
'metrics (target_obj={})').format(len(target_obj_idx),
target_objective))
metric_meta = MetricCatalog.objects.get_metric_meta(newest_result.session.dbms,
newest_result.session.target_objective)
if metric_meta[target_objective] == '(less is better)':
lessisbetter = True
else:
lessisbetter = False
y_workload = y_workload[:, target_obj_idx]
y_target = y_target[:, target_obj_idx]
y_columnlabels = y_columnlabels[target_obj_idx]
# Combine duplicate rows in the target/workload data (separately)
X_workload, y_workload, rowlabels_workload = DataUtil.combine_duplicate_rows(
X_workload, y_workload, rowlabels_workload)
X_target, y_target, rowlabels_target = DataUtil.combine_duplicate_rows(
X_target, y_target, rowlabels_target)
# Delete any rows that appear in both the workload data and the target
# data from the workload data
dups_filter = np.ones(X_workload.shape[0], dtype=bool)
target_row_tups = [tuple(row) for row in X_target]
for i, row in enumerate(X_workload):
if tuple(row) in target_row_tups:
dups_filter[i] = False
X_workload = X_workload[dups_filter, :]
y_workload = y_workload[dups_filter, :]
rowlabels_workload = rowlabels_workload[dups_filter]
# Combine target & workload Xs for preprocessing
X_matrix = np.vstack([X_target, X_workload])
# Dummy encode categorial variables
categorical_info = DataUtil.dummy_encoder_helper(X_columnlabels,
mapped_workload.dbms)
dummy_encoder = DummyEncoder(categorical_info['n_values'],
categorical_info['categorical_features'],
categorical_info['cat_columnlabels'],
categorical_info['noncat_columnlabels'])
X_matrix = dummy_encoder.fit_transform(X_matrix)
# below two variables are needed for correctly determing max/min on dummies
binary_index_set = set(categorical_info['binary_vars'])
total_dummies = dummy_encoder.total_dummies()
# Scale to N(0, 1)
X_scaler = StandardScaler()
X_scaled = X_scaler.fit_transform(X_matrix)
if y_target.shape[0] < 5: # FIXME
# FIXME (dva): if there are fewer than 5 target results so far
# then scale the y values (metrics) using the workload's
# y_scaler. I'm not sure if 5 is the right cutoff.
y_target_scaler = None
y_workload_scaler = StandardScaler()
y_matrix = np.vstack([y_target, y_workload])
y_scaled = y_workload_scaler.fit_transform(y_matrix)
else:
# FIXME (dva): otherwise try to compute a separate y_scaler for
# the target and scale them separately.
try:
y_target_scaler = StandardScaler()
y_workload_scaler = StandardScaler()
y_target_scaled = y_target_scaler.fit_transform(y_target)
y_workload_scaled = y_workload_scaler.fit_transform(y_workload)
y_scaled = np.vstack([y_target_scaled, y_workload_scaled])
except ValueError:
y_target_scaler = None
y_workload_scaler = StandardScaler()
y_scaled = y_workload_scaler.fit_transform(y_target)
# Set up constraint helper
constraint_helper = ParamConstraintHelper(scaler=X_scaler,
encoder=dummy_encoder,
binary_vars=categorical_info['binary_vars'],
init_flip_prob=INIT_FLIP_PROB,
flip_prob_decay=FLIP_PROB_DECAY)
# FIXME (dva): check if these are good values for the ridge
# ridge = np.empty(X_scaled.shape[0])
# ridge[:X_target.shape[0]] = 0.01
# ridge[X_target.shape[0]:] = 0.1
# FIXME: we should generate more samples and use a smarter sampling
# technique
num_samples = NUM_SAMPLES
X_samples = np.empty((num_samples, X_scaled.shape[1]))
X_min = np.empty(X_scaled.shape[1])
X_max = np.empty(X_scaled.shape[1])
knobs_mem = KnobCatalog.objects.filter(
dbms=newest_result.session.dbms, tunable=True, resource=1)
knobs_mem_catalog = {k.name: k for k in knobs_mem}
mem_max = newest_result.workload.hardware.memory
X_mem = np.zeros([1, X_scaled.shape[1]])
X_default = np.empty(X_scaled.shape[1])
# Get default knob values
for i, k_name in enumerate(X_columnlabels):
k = KnobCatalog.objects.filter(dbms=newest_result.session.dbms, name=k_name)[0]
X_default[i] = k.default
X_default_scaled = X_scaler.transform(X_default.reshape(1, X_default.shape[0]))[0]
# Determine min/max for knob values
for i in range(X_scaled.shape[1]):
if i < total_dummies or i in binary_index_set:
col_min = 0
col_max = 1
else:
col_min = X_scaled[:, i].min()
col_max = X_scaled[:, i].max()
if X_columnlabels[i] in knobs_mem_catalog:
X_mem[0][i] = mem_max * 1024 * 1024 * 1024 # mem_max GB
col_max = X_scaler.transform(X_mem)[0][i]
# Set min value to the default value
# FIXME: support multiple methods can be selected by users
col_min = X_default_scaled[i]
X_min[i] = col_min
X_max[i] = col_max
X_samples[:, i] = np.random.rand(num_samples) * (col_max - col_min) + col_min
# Maximize the throughput, moreisbetter
# Use gradient descent to minimize -throughput
if not lessisbetter:
y_scaled = -y_scaled
q = queue.PriorityQueue()
for x in range(0, y_scaled.shape[0]):
q.put((y_scaled[x][0], x))
i = 0
while i < TOP_NUM_CONFIG:
try:
item = q.get_nowait()
# Tensorflow get broken if we use the training data points as
# starting points for GPRGD. We add a small bias for the
# starting points. GPR_EPS default value is 0.001
X_samples = np.vstack((X_samples, X_scaled[item[1]] + GPR_EPS))
i = i + 1
except queue.Empty:
break
model = GPRGD(length_scale=DEFAULT_LENGTH_SCALE,
magnitude=DEFAULT_MAGNITUDE,
max_train_size=MAX_TRAIN_SIZE,
batch_size=BATCH_SIZE,
num_threads=NUM_THREADS,
learning_rate=DEFAULT_LEARNING_RATE,
epsilon=DEFAULT_EPSILON,
max_iter=MAX_ITER,
sigma_multiplier=DEFAULT_SIGMA_MULTIPLIER,
mu_multiplier=DEFAULT_MU_MULTIPLIER)
model.fit(X_scaled, y_scaled, X_min, X_max, ridge=DEFAULT_RIDGE)
res = model.predict(X_samples, constraint_helper=constraint_helper)
best_config_idx = np.argmin(res.minl.ravel())
best_config = res.minl_conf[best_config_idx, :]
best_config = X_scaler.inverse_transform(best_config)
# Decode one-hot encoding into categorical knobs
best_config = dummy_encoder.inverse_transform(best_config)
# Although we have max/min limits in the GPRGD training session, it may
# lose some precisions. e.g. 0.99..99 >= 1.0 may be True on the scaled data,
# when we inversely transform the scaled data, the different becomes much larger
# and cannot be ignored. Here we check the range on the original data
# directly, and make sure the recommended config lies within the range
X_min_inv = X_scaler.inverse_transform(X_min)
X_max_inv = X_scaler.inverse_transform(X_max)
best_config = np.minimum(best_config, X_max_inv)
best_config = np.maximum(best_config, X_min_inv)
conf_map = {k: best_config[i] for i, k in enumerate(X_columnlabels)}
conf_map_res = {}
conf_map_res['status'] = 'good'
conf_map_res['recommendation'] = conf_map
conf_map_res['info'] = 'INFO: training data size is {}'.format(X_scaled.shape[0])
return conf_map_res