def test_combine(self):
test_dedup_row_labels = np.array(["Workload-0", "Workload-1"])
test_dedup_x = np.matrix([[0.22, 5, "string", "11:11", "fsync", True],
[0.21, 6, "string", "11:12", "fsync", True]])
test_dedup_y = np.matrix([[30, 30, 40],
[10, 10, 40]])
test_x, test_y, row_labels = DataUtil.combine_duplicate_rows(
test_dedup_x, test_dedup_y, test_dedup_row_labels)
self.assertEqual(len(test_x), len(test_y))
self.assertEqual(len(test_x), len(row_labels))
self.assertEqual(row_labels[0], tuple([test_dedup_row_labels[0]]))
self.assertEqual(row_labels[1], tuple([test_dedup_row_labels[1]]))
self.assertTrue((test_x[0] == test_dedup_x[0]).all())
self.assertTrue((test_x[1] == test_dedup_x[1]).all())
self.assertTrue((test_y[0] == test_dedup_y[0]).all())
self.assertTrue((test_y[1] == test_dedup_y[1]).all())
test_row_labels = np.array(["Workload-0",
"Workload-1",
"Workload-2",
"Workload-3"])
test_x_matrix = np.matrix([[0.22, 5, "string", "timestamp", "enum", True],
[0.3, 5, "rstring", "timestamp2", "enum", False],
[0.22, 5, "string", "timestamp", "enum", True],
[0.3, 5, "r", "timestamp2", "enum", False]])
test_y_matrix = np.matrix([[20, 30, 40],
[30, 30, 40],
[20, 30, 40],
[32, 30, 40]])
test_x, test_y, row_labels = DataUtil.combine_duplicate_rows(
test_x_matrix, test_y_matrix, test_row_labels)
self.assertTrue(len(test_x) <= len(test_x_matrix))
self.assertTrue(len(test_y) <= len(test_y_matrix))
self.assertEqual(len(test_x), len(test_y))
self.assertEqual(len(test_x), len(row_labels))
row_labels_set = set(row_labels)
self.assertTrue(tuple(["Workload-0", "Workload-2"]) in row_labels_set)
self.assertTrue(("Workload-1",) in row_labels_set)
self.assertTrue(("Workload-3",) in row_labels_set)
rows = set()
for i in test_x:
self.assertTrue(tuple(i) not in rows)
self.assertTrue(i in test_x_matrix)
rows.add(tuple(i))
rowys = set()
for i in test_y:
self.assertTrue(tuple(i) not in rowys)
self.assertTrue(i in test_y_matrix)
rowys.add(tuple(i))
def test_combine(self):
test_dedup_row_labels = np.array(["Workload-0", "Workload-1"])
test_dedup_x = np.matrix([[0.22, 5, "string", "11:11", "fsync", True],
[0.21, 6, "string", "11:12", "fsync", True]])
test_dedup_y = np.matrix([[30, 30, 40],
[10, 10, 40]])
test_x, test_y, row_labels = DataUtil.combine_duplicate_rows(
test_dedup_x, test_dedup_y, test_dedup_row_labels)
self.assertEqual(len(test_x), len(test_y))
self.assertEqual(len(test_x), len(row_labels))
self.assertEqual(row_labels[0], tuple([test_dedup_row_labels[0]]))
self.assertEqual(row_labels[1], tuple([test_dedup_row_labels[1]]))
self.assertTrue((test_x[0] == test_dedup_x[0]).all())
self.assertTrue((test_x[1] == test_dedup_x[1]).all())
self.assertTrue((test_y[0] == test_dedup_y[0]).all())
self.assertTrue((test_y[1] == test_dedup_y[1]).all())
test_row_labels = np.array(["Workload-0",
"Workload-1",
"Workload-2",
"Workload-3"])
test_x_matrix = np.matrix([[0.22, 5, "string", "timestamp", "enum", True],
[0.3, 5, "rstring", "timestamp2", "enum", False],
[0.22, 5, "string", "timestamp", "enum", True],
[0.3, 5, "r", "timestamp2", "enum", False]])
test_y_matrix = np.matrix([[20, 30, 40],
[30, 30, 40],
[20, 30, 40],
[32, 30, 40]])
test_x, test_y, row_labels = DataUtil.combine_duplicate_rows(
test_x_matrix, test_y_matrix, test_row_labels)
self.assertTrue(len(test_x) <= len(test_x_matrix))
self.assertTrue(len(test_y) <= len(test_y_matrix))
self.assertEqual(len(test_x), len(test_y))
self.assertEqual(len(test_x), len(row_labels))
row_labels_set = set(row_labels)
self.assertTrue(tuple(["Workload-0", "Workload-2"]) in row_labels_set)
self.assertTrue(("Workload-1",) in row_labels_set)
self.assertTrue(("Workload-3",) in row_labels_set)
rows = set()
for i in test_x:
self.assertTrue(tuple(i) not in rows)
self.assertTrue(i in test_x_matrix)
rows.add(tuple(i))
rowys = set()
for i in test_y:
self.assertTrue(tuple(i) not in rowys)
self.assertTrue(i in test_y_matrix)
rowys.add(tuple(i))
def map_workload(target_data):
# Get the latest version of pipeline data that's been computed so far.
latest_pipeline_run = PipelineRun.objects.get_latest()
if target_data['bad']:
assert target_data is not None
return target_data
assert latest_pipeline_run is not None
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
target_workload = newest_result.workload
X_columnlabels = np.array(target_data['X_columnlabels'])
y_columnlabels = np.array(target_data['y_columnlabels'])
# Find all pipeline data belonging to the latest version with the same
# DBMS and hardware as the target
pipeline_data = PipelineData.objects.filter(
pipeline_run=latest_pipeline_run,
workload__dbms=target_workload.dbms,
workload__hardware=target_workload.hardware)
# FIXME (dva): we should also compute the global (i.e., overall) ranked_knobs
# and pruned metrics but we just use those from the first workload for now
initialized = False
global_ranked_knobs = None
global_pruned_metrics = None
ranked_knob_idxs = None
pruned_metric_idxs = None
# Compute workload mapping data for each unique workload
unique_workloads = pipeline_data.values_list('workload',
flat=True).distinct()
assert len(unique_workloads) > 0
workload_data = {}
for unique_workload in unique_workloads:
workload_obj = Workload.objects.get(pk=unique_workload)
wkld_results = Result.objects.filter(workload=workload_obj)
if wkld_results.exists() is False:
# delete the workload
workload_obj.delete()
continue
# Load knob & metric data for this workload
knob_data = load_data_helper(pipeline_data, unique_workload,
PipelineTaskType.KNOB_DATA)
metric_data = load_data_helper(pipeline_data, unique_workload,
PipelineTaskType.METRIC_DATA)
X_matrix = np.array(knob_data["data"])
y_matrix = np.array(metric_data["data"])
rowlabels = np.array(knob_data["rowlabels"])
assert np.array_equal(rowlabels, metric_data["rowlabels"])
if not initialized:
# For now set ranked knobs & pruned metrics to be those computed
# for the first workload
global_ranked_knobs = load_data_helper(
pipeline_data, unique_workload,
PipelineTaskType.RANKED_KNOBS)[:IMPORTANT_KNOB_NUMBER]
global_pruned_metrics = load_data_helper(
pipeline_data, unique_workload,
PipelineTaskType.PRUNED_METRICS)
ranked_knob_idxs = [
i for i in range(X_matrix.shape[1])
if X_columnlabels[i] in global_ranked_knobs
]
pruned_metric_idxs = [
i for i in range(y_matrix.shape[1])
if y_columnlabels[i] in global_pruned_metrics
]
# Filter X & y columnlabels by top ranked_knobs & pruned_metrics
X_columnlabels = X_columnlabels[ranked_knob_idxs]
y_columnlabels = y_columnlabels[pruned_metric_idxs]
initialized = True
# Filter X & y matrices by top ranked_knobs & pruned_metrics
X_matrix = X_matrix[:, ranked_knob_idxs]
y_matrix = y_matrix[:, pruned_metric_idxs]
# Combine duplicate rows (rows with same knob settings)
X_matrix, y_matrix, rowlabels = DataUtil.combine_duplicate_rows(
X_matrix, y_matrix, rowlabels)
workload_data[unique_workload] = {
'X_matrix': X_matrix,
'y_matrix': y_matrix,
'rowlabels': rowlabels,
}
assert len(workload_data) > 0
# Stack all X & y matrices for preprocessing
Xs = np.vstack(
[entry['X_matrix'] for entry in list(workload_data.values())])
ys = np.vstack(
[entry['y_matrix'] for entry in list(workload_data.values())])
# Scale the X & y values, then compute the deciles for each column in y
X_scaler = StandardScaler(copy=False)
X_scaler.fit(Xs)
y_scaler = StandardScaler(copy=False)
y_scaler.fit_transform(ys)
y_binner = Bin(bin_start=1, axis=0)
y_binner.fit(ys)
del Xs
del ys
# Filter the target's X & y data by the ranked knobs & pruned metrics.
X_target = target_data['X_matrix'][:, ranked_knob_idxs]
y_target = target_data['y_matrix'][:, pruned_metric_idxs]
# Now standardize the target's data and bin it by the deciles we just
# calculated
X_target = X_scaler.transform(X_target)
y_target = y_scaler.transform(y_target)
y_target = y_binner.transform(y_target)
scores = {}
for workload_id, workload_entry in list(workload_data.items()):
predictions = np.empty_like(y_target)
X_workload = workload_entry['X_matrix']
X_scaled = X_scaler.transform(X_workload)
y_workload = workload_entry['y_matrix']
y_scaled = y_scaler.transform(y_workload)
for j, y_col in enumerate(y_scaled.T):
# Using this workload's data, train a Gaussian process model
# and then predict the performance of each metric for each of
# the knob configurations attempted so far by the target.
y_col = y_col.reshape(-1, 1)
model = GPRNP(length_scale=DEFAULT_LENGTH_SCALE,
magnitude=DEFAULT_MAGNITUDE,
max_train_size=MAX_TRAIN_SIZE,
batch_size=BATCH_SIZE)
model.fit(X_scaled, y_col, ridge=DEFAULT_RIDGE)
predictions[:, j] = model.predict(X_target).ypreds.ravel()
# Bin each of the predicted metric columns by deciles and then
# compute the score (i.e., distance) between the target workload
# and each of the known workloads
predictions = y_binner.transform(predictions)
dists = np.sqrt(
np.sum(np.square(np.subtract(predictions, y_target)), axis=1))
scores[workload_id] = np.mean(dists)
# Find the best (minimum) score
best_score = np.inf
best_workload_id = None
# scores_info = {workload_id: (workload_name, score)}
scores_info = {}
for workload_id, similarity_score in list(scores.items()):
workload_name = Workload.objects.get(pk=workload_id).name
if similarity_score < best_score:
best_score = similarity_score
best_workload_id = workload_id
best_workload_name = workload_name
scores_info[workload_id] = (workload_name, similarity_score)
target_data['mapped_workload'] = (best_workload_id, best_workload_name,
best_score)
target_data['scores'] = scores_info
return target_data
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 combine_workload(target_data):
# 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'])
session = newest_result.session
params = JSONUtil.loads(session.hyperparameters)
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)'),
X_columnlabels, target_data['X_columnlabels'])
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)'),
y_columnlabels, target_data['y_columnlabels'])
# 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)[:params['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))
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
if ENABLE_DUMMY_ENCODER:
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)
binary_encoder = categorical_info['binary_vars']
# 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()
else:
dummy_encoder = None
binary_encoder = None
binary_index_set = []
total_dummies = 0
# 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)
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
# Maximize the throughput, moreisbetter
# Use gradient descent to minimize -throughput
if not lessisbetter:
y_scaled = -y_scaled
# Set up constraint helper
constraint_helper = ParamConstraintHelper(
scaler=X_scaler,
encoder=dummy_encoder,
binary_vars=binary_encoder,
init_flip_prob=params['INIT_FLIP_PROB'],
flip_prob_decay=params['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
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
return X_columnlabels, X_scaler, X_scaled, y_scaled, X_max, X_min,\
dummy_encoder, constraint_helper
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)[:10] # FIXME
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))
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 then scale
X_matrix = np.vstack([X_target, X_workload])
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)
# 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 = 20
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])
for i in range(X_scaled.shape[1]):
col_min = X_scaled[:, i].min()
col_max = X_scaled[:, i].max()
X_min[i] = col_min
X_max[i] = col_max
X_samples[:, i] = np.random.rand(
num_samples) * (col_max - col_min) + col_min
# FIXME: Maximize the throughput, hardcode
# Use gradient descent to minimize -throughput
y_scaled = -y_scaled
model = GPRGD()
model.fit(X_scaled, y_scaled, X_min, X_max, ridge)
res = model.predict(X_samples)
# FIXME: whether we select the min/max for the best config depends
# on the target objective
best_config_idx = np.argmin(res.minl.ravel())
best_config = res.minl_conf[best_config_idx, :]
best_config = X_scaler.inverse_transform(best_config)
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 map_workload(map_workload_input):
start_ts = time.time()
target_data, algorithm = map_workload_input
if target_data['bad']:
assert target_data is not None
target_data['pipeline_run'] = None
LOG.debug('%s: Skipping workload mapping.\n\ndata=%s\n',
AlgorithmType.name(algorithm),
JSONUtil.dumps(target_data, pprint=True))
return target_data, algorithm
# Get the latest version of pipeline data that's been computed so far.
latest_pipeline_run = PipelineRun.objects.get_latest()
assert latest_pipeline_run is not None
target_data['pipeline_run'] = latest_pipeline_run.pk
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
session = newest_result.session
params = JSONUtil.loads(session.hyperparameters)
target_workload = newest_result.workload
X_columnlabels = np.array(target_data['X_columnlabels'])
y_columnlabels = np.array(target_data['y_columnlabels'])
# Find all pipeline data belonging to the latest version with the same
# DBMS and hardware as the target
pipeline_data = PipelineData.objects.filter(
pipeline_run=latest_pipeline_run,
workload__dbms=target_workload.dbms,
workload__hardware=target_workload.hardware,
workload__project=target_workload.project)
# FIXME (dva): we should also compute the global (i.e., overall) ranked_knobs
# and pruned metrics but we just use those from the first workload for now
initialized = False
global_ranked_knobs = None
global_pruned_metrics = None
ranked_knob_idxs = None
pruned_metric_idxs = None
unique_workloads = pipeline_data.values_list('workload',
flat=True).distinct()
workload_data = {}
# Compute workload mapping data for each unique workload
for unique_workload in unique_workloads:
workload_obj = Workload.objects.get(pk=unique_workload)
wkld_results = Result.objects.filter(workload=workload_obj)
if wkld_results.exists() is False:
# delete the workload
workload_obj.delete()
continue
# Load knob & metric data for this workload
knob_data = load_data_helper(pipeline_data, unique_workload,
PipelineTaskType.KNOB_DATA)
knob_data["data"], knob_data["columnlabels"] = clean_knob_data(
knob_data["data"], knob_data["columnlabels"],
newest_result.session)
metric_data = load_data_helper(pipeline_data, unique_workload,
PipelineTaskType.METRIC_DATA)
X_matrix = np.array(knob_data["data"])
y_matrix = np.array(metric_data["data"])
rowlabels = np.array(knob_data["rowlabels"])
assert np.array_equal(rowlabels, metric_data["rowlabels"])
if not initialized:
# For now set ranked knobs & pruned metrics to be those computed
# for the first workload
global_ranked_knobs = load_data_helper(
pipeline_data, unique_workload, PipelineTaskType.RANKED_KNOBS
)[:params['IMPORTANT_KNOB_NUMBER']]
global_pruned_metrics = load_data_helper(
pipeline_data, unique_workload,
PipelineTaskType.PRUNED_METRICS)
ranked_knob_idxs = [
i for i in range(X_matrix.shape[1])
if X_columnlabels[i] in global_ranked_knobs
]
pruned_metric_idxs = [
i for i in range(y_matrix.shape[1])
if y_columnlabels[i] in global_pruned_metrics
]
# Filter X & y columnlabels by top ranked_knobs & pruned_metrics
X_columnlabels = X_columnlabels[ranked_knob_idxs]
y_columnlabels = y_columnlabels[pruned_metric_idxs]
initialized = True
# Filter X & y matrices by top ranked_knobs & pruned_metrics
X_matrix = X_matrix[:, ranked_knob_idxs]
y_matrix = y_matrix[:, pruned_metric_idxs]
# Combine duplicate rows (rows with same knob settings)
X_matrix, y_matrix, rowlabels = DataUtil.combine_duplicate_rows(
X_matrix, y_matrix, rowlabels)
workload_data[unique_workload] = {
'X_matrix': X_matrix,
'y_matrix': y_matrix,
'rowlabels': rowlabels,
}
if len(workload_data) == 0:
# The background task that aggregates the data has not finished running yet
target_data.update(mapped_workload=None, scores=None)
LOG.debug(
'%s: Skipping workload mapping because there is no parsed workload.\n',
AlgorithmType.name(algorithm))
return target_data, algorithm
# Stack all X & y matrices for preprocessing
Xs = np.vstack(
[entry['X_matrix'] for entry in list(workload_data.values())])
ys = np.vstack(
[entry['y_matrix'] for entry in list(workload_data.values())])
# Scale the X & y values, then compute the deciles for each column in y
X_scaler = StandardScaler(copy=False)
X_scaler.fit(Xs)
y_scaler = StandardScaler(copy=False)
y_scaler.fit_transform(ys)
y_binner = Bin(bin_start=1, axis=0)
y_binner.fit(ys)
del Xs
del ys
# Filter the target's X & y data by the ranked knobs & pruned metrics.
X_target = target_data['X_matrix'][:, ranked_knob_idxs]
y_target = target_data['y_matrix'][:, pruned_metric_idxs]
# Now standardize the target's data and bin it by the deciles we just
# calculated
X_target = X_scaler.transform(X_target)
y_target = y_scaler.transform(y_target)
y_target = y_binner.transform(y_target)
scores = {}
for workload_id, workload_entry in list(workload_data.items()):
predictions = np.empty_like(y_target)
X_workload = workload_entry['X_matrix']
X_scaled = X_scaler.transform(X_workload)
y_workload = workload_entry['y_matrix']
y_scaled = y_scaler.transform(y_workload)
for j, y_col in enumerate(y_scaled.T):
# Using this workload's data, train a Gaussian process model
# and then predict the performance of each metric for each of
# the knob configurations attempted so far by the target.
y_col = y_col.reshape(-1, 1)
if params['GPR_USE_GPFLOW']:
model_kwargs = {
'lengthscales': params['GPR_LENGTH_SCALE'],
'variance': params['GPR_MAGNITUDE'],
'noise_variance': params['GPR_RIDGE']
}
tf.reset_default_graph()
graph = tf.get_default_graph()
gpflow.reset_default_session(graph=graph)
m = gpr_models.create_model(params['GPR_MODEL_NAME'],
X=X_scaled,
y=y_col,
**model_kwargs)
gpr_result = gpflow_predict(m.model, X_target)
else:
model = GPRNP(length_scale=params['GPR_LENGTH_SCALE'],
magnitude=params['GPR_MAGNITUDE'],
max_train_size=params['GPR_MAX_TRAIN_SIZE'],
batch_size=params['GPR_BATCH_SIZE'])
model.fit(X_scaled, y_col, ridge=params['GPR_RIDGE'])
gpr_result = model.predict(X_target)
predictions[:, j] = gpr_result.ypreds.ravel()
# Bin each of the predicted metric columns by deciles and then
# compute the score (i.e., distance) between the target workload
# and each of the known workloads
predictions = y_binner.transform(predictions)
dists = np.sqrt(
np.sum(np.square(np.subtract(predictions, y_target)), axis=1))
scores[workload_id] = np.mean(dists)
# Find the best (minimum) score
best_score = np.inf
best_workload_id = None
best_workload_name = None
scores_info = {}
for workload_id, similarity_score in list(scores.items()):
workload_name = Workload.objects.get(pk=workload_id).name
if similarity_score < best_score:
best_score = similarity_score
best_workload_id = workload_id
best_workload_name = workload_name
scores_info[workload_id] = (workload_name, similarity_score)
target_data.update(mapped_workload=(best_workload_id, best_workload_name,
best_score),
scores=scores_info)
LOG.debug('%s: Finished mapping the workload.\n\ndata=%s\n',
AlgorithmType.name(algorithm),
JSONUtil.dumps(target_data, pprint=True))
save_execution_time(start_ts, "map_workload", newest_result)
return target_data, algorithm
def create_workload_mapping_data():
agg_datas = PipelineResult.objects.filter(
task_type=PipelineTaskType.AGGREGATED_DATA)
dbmss = set([ad.dbms.pk for ad in agg_datas])
hardwares = set([ad.hardware.pk for ad in agg_datas])
for dbms_id, hw_id in itertools.product(dbmss, hardwares):
data = PipelineResult.get_latest(dbms_id, hw_id,
PipelineTaskType.AGGREGATED_DATA)
file_info = JSONUtil.loads(data.value)
cluster_data = OrderedDict()
for cluster, path in file_info['data'].iteritems():
compressed_data = np.load(path)
X_matrix = compressed_data['X_matrix']
y_matrix = compressed_data['y_matrix']
X_columnlabels = compressed_data['X_columnlabels']
y_columnlabels = compressed_data['y_columnlabels']
rowlabels = compressed_data['rowlabels']
# Filter metrics and knobs
ranked_knobs = JSONUtil.loads(
PipelineResult.get_latest(
dbms_id, hw_id,
PipelineTaskType.RANKED_KNOBS).value)[:10] # FIXME
pruned_metrics = JSONUtil.loads(
PipelineResult.get_latest(
dbms_id, hw_id, PipelineTaskType.PRUNED_METRICS).value)
knob_idxs = [
i for i in range(X_matrix.shape[1])
if X_columnlabels[i] in ranked_knobs
]
metric_idxs = [
i for i in range(y_matrix.shape[1])
if y_columnlabels[i] in pruned_metrics
]
X_matrix = X_matrix[:, knob_idxs]
X_columnlabels = X_columnlabels[knob_idxs]
y_matrix = y_matrix[:, metric_idxs]
y_columnlabels = y_columnlabels[metric_idxs]
# Combine duplicate rows
X_matrix, y_matrix, rowlabels = DataUtil.combine_duplicate_rows(
X_matrix, y_matrix, rowlabels)
cluster_data[cluster] = {
'X_matrix': X_matrix,
'y_matrix': y_matrix,
'X_columnlabels': X_columnlabels,
'y_columnlabels': y_columnlabels,
'rowlabels': rowlabels,
}
Xs = np.vstack([entry['X_matrix'] for entry in cluster_data.values()])
ys = np.vstack([entry['y_matrix'] for entry in cluster_data.values()])
X_scaler = StandardScaler(copy=False)
X_scaler.fit(Xs)
y_scaler = StandardScaler(copy=False)
y_scaler.fit_transform(ys)
y_binner = Bin(axis=0)
y_binner.fit(ys)
del Xs
del ys
task_name = PipelineTaskType.TYPE_NAMES[
PipelineTaskType.WORKLOAD_MAPPING_DATA].replace(' ', '').upper()
timestamp = data.creation_timestamp
tsf = timestamp.strftime("%Y%m%d-%H%M%S")
savepaths = {}
for cluster, entry in cluster_data.iteritems():
X_scaler.transform(entry['X_matrix'])
y_scaler.transform(entry['y_matrix'])
fname = '{}_{}_{}_{}_{}.npz'.format(task_name, dbms_id, hw_id,
cluster, tsf)
savepath = os.path.join(PIPELINE_DIR, fname)
savepaths[cluster] = savepath
np.savez_compressed(savepath, **entry)
X_scaler_path = os.path.join(
PIPELINE_DIR,
'{}_XSCALER_{}_{}_{}.npz'.format(task_name, dbms_id, hw_id, tsf))
np.savez_compressed(X_scaler_path,
mean=X_scaler.mean_,
scale=X_scaler.scale_)
y_scaler_path = os.path.join(
PIPELINE_DIR,
'{}_YSCALER_{}_{}_{}.npz'.format(task_name, dbms_id, hw_id, tsf))
np.savez_compressed(y_scaler_path,
mean=y_scaler.mean_,
scale=y_scaler.scale_)
y_deciles_path = os.path.join(
PIPELINE_DIR,
'{}_YDECILES_{}_{}_{}.npz'.format(task_name, dbms_id, hw_id, tsf))
np.savez_compressed(y_deciles_path, deciles=y_binner.deciles_)
value = {
'data': savepaths,
'X_scaler': X_scaler_path,
'y_scaler': y_scaler_path,
'y_deciles': y_deciles_path,
'X_columnlabels':
cluster_data.values()[0]['X_columnlabels'].tolist(),
'y_columnlabels':
cluster_data.values()[0]['y_columnlabels'].tolist(),
}
new_res = PipelineResult()
new_res.dbms = DBMSCatalog.objects.get(pk=dbms_id)
new_res.hardware = Hardware.objects.get(pk=hw_id)
new_res.creation_timestamp = timestamp
new_res.task_type = PipelineTaskType.WORKLOAD_MAPPING_DATA
new_res.value = JSONUtil.dumps(value, pprint=True)
new_res.save()
def configuration_recommendation(target_data):
if target_data['scores'] is None:
raise NotImplementedError('Implement me!')
best_wkld_id = target_data['mapped_workload'][0]
# Load specific workload data
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
target_obj = newest_result.application.target_objective
dbms_id = newest_result.dbms.pk
hw_id = newest_result.application.hardware.pk
agg_data = PipelineResult.get_latest(dbms_id, hw_id,
PipelineTaskType.AGGREGATED_DATA)
if agg_data is None:
return None
data_map = JSONUtil.loads(agg_data.value)
if best_wkld_id not in data_map['data']:
raise Exception(('Cannot find mapped workload'
'(id={}) in aggregated data').format(best_wkld_id))
workload_data = np.load(data_map['data'][best_wkld_id])
# Mapped workload data
X_wkld_matrix = workload_data['X_matrix']
y_wkld_matrix = workload_data['y_matrix']
wkld_rowlabels = workload_data['rowlabels']
X_columnlabels = workload_data['X_columnlabels']
y_columnlabels = workload_data['y_columnlabels']
# Target workload data
X_target_matrix = target_data['X_matrix']
y_target_matrix = target_data['y_matrix']
target_rowlabels = 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 knobs
ranked_knobs = JSONUtil.loads(
PipelineResult.get_latest(
dbms_id, hw_id, PipelineTaskType.RANKED_KNOBS).value)[:10] # FIXME
X_idxs = [
i for i in range(X_columnlabels.shape[0])
if X_columnlabels[i] in ranked_knobs
]
X_wkld_matrix = X_wkld_matrix[:, X_idxs]
X_target_matrix = X_target_matrix[:, X_idxs]
X_columnlabels = X_columnlabels[X_idxs]
# Filter metrics by current target objective metric
y_idx = [
i for i in range(y_columnlabels.shape[0])
if y_columnlabels[i] == target_obj
]
if len(y_idx) == 0:
raise Exception(('Could not find target objective in metrics '
'(target_obj={})').format(target_obj))
elif len(y_idx) > 1:
raise Exception(
('Found {} instances of target objective in '
'metrics (target_obj={})').format(len(y_idx), target_obj))
y_wkld_matrix = y_wkld_matrix[:, y_idx]
y_target_matrix = y_target_matrix[:, y_idx]
y_columnlabels = y_columnlabels[y_idx]
# Combine duplicate rows in the target/workload data (separately)
X_wkld_matrix, y_wkld_matrix, wkld_rowlabels = DataUtil.combine_duplicate_rows(
X_wkld_matrix, y_wkld_matrix, wkld_rowlabels)
X_target_matrix, y_target_matrix, target_rowlabels = DataUtil.combine_duplicate_rows(
X_target_matrix, y_target_matrix, target_rowlabels)
# Delete any rows that appear in both the workload data and the target
# data from the workload data
dups_filter = np.ones(X_wkld_matrix.shape[0], dtype=bool)
target_row_tups = [tuple(row) for row in X_target_matrix]
for i, row in enumerate(X_wkld_matrix):
if tuple(row) in target_row_tups:
dups_filter[i] = False
X_wkld_matrix = X_wkld_matrix[dups_filter, :]
y_wkld_matrix = y_wkld_matrix[dups_filter, :]
wkld_rowlabels = wkld_rowlabels[dups_filter]
# Combine Xs and scale
X_matrix = np.vstack([X_target_matrix, X_wkld_matrix])
X_scaler = StandardScaler()
X_scaled = X_scaler.fit_transform(X_matrix)
if y_target_matrix.shape[0] < 5: # FIXME
y_target_scaler = None
y_wkld_scaler = StandardScaler()
y_matrix = np.vstack([y_target_matrix, y_wkld_matrix])
y_scaled = y_wkld_scaler.fit_transform(y_matrix)
else:
try:
y_target_scaler = StandardScaler()
y_wkld_scaler = StandardScaler()
y_target_scaled = y_target_scaler.fit_transform(y_target_matrix)
y_wkld_scaled = y_wkld_scaler.fit_transform(y_wkld_matrix)
y_scaled = np.vstack([y_target_scaled, y_wkld_scaled])
except ValueError:
y_target_scaler = None
y_wkld_scaler = StandardScaler()
y_matrix = np.vstack([y_target_matrix, y_wkld_matrix])
y_scaled = y_wkld_scaler.fit_transform(y_matrix)
ridge = np.empty(X_scaled.shape[0])
ridge[:X_target_matrix.shape[0]] = 0.01
ridge[X_target_matrix.shape[0]:] = 0.1
# FIXME
num_samples = 5
X_samples = np.empty((num_samples, X_scaled.shape[1]))
for i in range(X_scaled.shape[1]):
col_min = X_scaled[:, i].min()
col_max = X_scaled[:, i].max()
X_samples[:, i] = np.random.rand(num_samples) * (col_max -
col_min) + col_min
model = GPR_GD()
model.fit(X_scaled, y_scaled, ridge)
res = model.predict(X_samples)
best_idx = np.argmin(res.minL.ravel())
best_conf = res.minL_conf[best_idx, :]
best_conf = X_scaler.inverse_transform(best_conf)
conf_map = {k: best_conf[i] for i, k in enumerate(X_columnlabels)}
return conf_map
def map_workload(target_data):
# Get the latest version of pipeline data that's been computed so far.
latest_pipeline_run = PipelineRun.objects.get_latest()
if target_data['bad']:
assert target_data is not None
return target_data
assert latest_pipeline_run is not None
newest_result = Result.objects.get(pk=target_data['newest_result_id'])
target_workload = newest_result.workload
X_columnlabels = np.array(target_data['X_columnlabels'])
y_columnlabels = np.array(target_data['y_columnlabels'])
# Find all pipeline data belonging to the latest version with the same
# DBMS and hardware as the target
pipeline_data = PipelineData.objects.filter(
pipeline_run=latest_pipeline_run,
workload__dbms=target_workload.dbms,
workload__hardware=target_workload.hardware)
# FIXME (dva): we should also compute the global (i.e., overall) ranked_knobs
# and pruned metrics but we just use those from the first workload for now
initialized = False
global_ranked_knobs = None
global_pruned_metrics = None
ranked_knob_idxs = None
pruned_metric_idxs = None
# Compute workload mapping data for each unique workload
unique_workloads = pipeline_data.values_list('workload', flat=True).distinct()
assert len(unique_workloads) > 0
workload_data = {}
for unique_workload in unique_workloads:
# Load knob & metric data for this workload
knob_data = load_data_helper(pipeline_data, unique_workload, PipelineTaskType.KNOB_DATA)
metric_data = load_data_helper(pipeline_data, unique_workload, PipelineTaskType.METRIC_DATA)
X_matrix = np.array(knob_data["data"])
y_matrix = np.array(metric_data["data"])
rowlabels = np.array(knob_data["rowlabels"])
assert np.array_equal(rowlabels, metric_data["rowlabels"])
if not initialized:
# For now set ranked knobs & pruned metrics to be those computed
# for the first workload
global_ranked_knobs = load_data_helper(
pipeline_data, unique_workload,
PipelineTaskType.RANKED_KNOBS)[:IMPORTANT_KNOB_NUMBER]
global_pruned_metrics = load_data_helper(
pipeline_data, unique_workload, PipelineTaskType.PRUNED_METRICS)
ranked_knob_idxs = [i for i in range(X_matrix.shape[1]) if X_columnlabels[
i] in global_ranked_knobs]
pruned_metric_idxs = [i for i in range(y_matrix.shape[1]) if y_columnlabels[
i] in global_pruned_metrics]
# Filter X & y columnlabels by top ranked_knobs & pruned_metrics
X_columnlabels = X_columnlabels[ranked_knob_idxs]
y_columnlabels = y_columnlabels[pruned_metric_idxs]
initialized = True
# Filter X & y matrices by top ranked_knobs & pruned_metrics
X_matrix = X_matrix[:, ranked_knob_idxs]
y_matrix = y_matrix[:, pruned_metric_idxs]
# Combine duplicate rows (rows with same knob settings)
X_matrix, y_matrix, rowlabels = DataUtil.combine_duplicate_rows(
X_matrix, y_matrix, rowlabels)
workload_data[unique_workload] = {
'X_matrix': X_matrix,
'y_matrix': y_matrix,
'rowlabels': rowlabels,
}
# Stack all X & y matrices for preprocessing
Xs = np.vstack([entry['X_matrix'] for entry in list(workload_data.values())])
ys = np.vstack([entry['y_matrix'] for entry in list(workload_data.values())])
# Scale the X & y values, then compute the deciles for each column in y
X_scaler = StandardScaler(copy=False)
X_scaler.fit(Xs)
y_scaler = StandardScaler(copy=False)
y_scaler.fit_transform(ys)
y_binner = Bin(bin_start=1, axis=0)
y_binner.fit(ys)
del Xs
del ys
# Filter the target's X & y data by the ranked knobs & pruned metrics.
X_target = target_data['X_matrix'][:, ranked_knob_idxs]
y_target = target_data['y_matrix'][:, pruned_metric_idxs]
# Now standardize the target's data and bin it by the deciles we just
# calculated
X_target = X_scaler.transform(X_target)
y_target = y_scaler.transform(y_target)
y_target = y_binner.transform(y_target)
scores = {}
for workload_id, workload_entry in list(workload_data.items()):
predictions = np.empty_like(y_target)
X_workload = workload_entry['X_matrix']
X_scaled = X_scaler.transform(X_workload)
y_workload = workload_entry['y_matrix']
y_scaled = y_scaler.transform(y_workload)
for j, y_col in enumerate(y_scaled.T):
# Using this workload's data, train a Gaussian process model
# and then predict the performance of each metric for each of
# the knob configurations attempted so far by the target.
y_col = y_col.reshape(-1, 1)
model = GPRNP(length_scale=DEFAULT_LENGTH_SCALE,
magnitude=DEFAULT_MAGNITUDE,
max_train_size=MAX_TRAIN_SIZE,
batch_size=BATCH_SIZE)
model.fit(X_scaled, y_col, ridge=DEFAULT_RIDGE)
predictions[:, j] = model.predict(X_target).ypreds.ravel()
# Bin each of the predicted metric columns by deciles and then
# compute the score (i.e., distance) between the target workload
# and each of the known workloads
predictions = y_binner.transform(predictions)
dists = np.sqrt(np.sum(np.square(
np.subtract(predictions, y_target)), axis=1))
scores[workload_id] = np.mean(dists)
# Find the best (minimum) score
best_score = np.inf
best_workload_id = None
for workload_id, similarity_score in list(scores.items()):
if similarity_score < best_score:
best_score = similarity_score
best_workload_id = workload_id
target_data['mapped_workload'] = (best_workload_id, best_score)
target_data['scores'] = scores
return target_data
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
