def __init__(self, dump_pth=None, ordinal=True):
conf = LumosConf()
self.ds_root_pth = conf.get('dataset', 'path')
self.vendor_cnt = conf.get('dataset', 'vendor_cnt')
self.__data = None
self.dump_pth = dump_pth
# sampling interval
self.sampling_interval = 5
# the label is ordinal or raw
self.ordinal = ordinal
class DataLoader(object):
'''
Load training or testing data
'''
def __init__(self, no_dup=False, dump_pth=None):
self.conf = LumosConf()
self.no_dup = no_dup
self.ds_root_pth = self.conf.get('dataset', 'path')
self.vendor_cnt = self.conf.get('dataset', 'vendor_cnt')
self.__data = defaultdict(lambda: defaultdict(lambda: []))
self.dump_pth = dump_pth
def load_data(self):
if self.dump_pth:
self.__load_data_from_file()
return
def is_vendor(v):
return '.' not in v
no_dup_set = set()
for vendor in os.listdir(self.ds_root_pth):
if not is_vendor(vendor): continue
pth1 = os.path.join(self.ds_root_pth, vendor)
for inst_type in os.listdir(pth1):
pth2 = os.path.join(pth1, inst_type)
for w in os.listdir(pth2):
[scale, rnd] = w.strip().split('_')[-2:]
workload = '_'.join(w.strip().split('_')[:2])
if self.no_dup:
w_key = '_'.join((inst_type, workload, scale))
if w_key in no_dup_set: continue
no_dup_set.add(w_key)
pth3 = os.path.join(pth2, w)
pth_report = os.path.join(pth3, 'report.json')
pth_metrics = os.path.join(pth3, 'sar.csv')
[ts, jct] = mget_json_values(pth_report, 'timestamp',
'elapsed_time')
ts = encode_timestamp(ts)
jct = float(jct)
header, metrics = read_csv(pth_metrics)
if not header or not metrics: continue
norm_metrics = normalize_metrics(metrics)
self.__data[workload][vendor].append(
RecordEntry(inst_type, scale, norm_metrics, jct, ts))
def __load_data_from_file(self):
with open(self.dump_pth, 'rb') as fd:
self.__data = dill.load(fd)
def get_data(self):
return self.__data
def __stat(self, raw_series, idx):
conf = LumosConf()
selected_idxes = conf.get('dataset', 'selected_idx')
valid_max_val = conf.get_global_max_val(selected_idxes[idx])
max_val = np.max(raw_series) / valid_max_val
min_val = np.min(raw_series) / valid_max_val
avg_val = np.mean(raw_series) / valid_max_val
# var_val = np.var(raw_series) / valid_max_val
return [max_val, min_val, avg_val]
def get_train_test_data_external(test_wl,
train_scale,
truncate=False,
ordinal=False):
assert train_scale == 'small', 'currently the model evaluated using small as the trianing scale'
conf = LumosConf()
dmp_pre = conf.get('dataset', 'train_test_dump_prefix')
dmp_suf = 'o%d_t%d' % (ordinal, truncate)
wl_pth = os.path.join(dmp_pre, '%s_%s.pkl' % (test_wl, dmp_suf))
with open(wl_pth, 'rb') as fd:
(train_data, test_data) = dill.load(fd)
return train_data, test_data
def encode(self, norm_data, raw_data, sampling_interval=5):
ret = []
conf = LumosConf()
valid_idx = conf.get('dataset', 'selected_idx')
if norm_data.shape[1] != len(valid_idx):
norm_data = norm_data[:, valid_idx]
if raw_data.shape[1] != len(valid_idx):
raw_data = raw_data[:, valid_idx]
for i in range(norm_data.shape[1]):
tmp = []
norm_series = norm_data[:, i]
raw_series = raw_data[:, i]
fft_feat = self.__fft(norm_series,
sampling_interval=sampling_interval)
stat_feat = self.__stat(raw_series, i)
tmp.extend(fft_feat)
tmp.extend(stat_feat)
ret.extend(tmp)
return ret
abs_err = optimal_bar - optimal
rel_err = abs_err / optimal
err[scale]['abs_err'] = abs_err
err[scale]['rel_err'] = rel_err
avg_abs_err = np.mean(err['large']['abs_err'])
avg_rel_err = np.mean(err['large']['rel_err'])
return avg_abs_err, avg_rel_err
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='grid search')
parser.add_argument('-j', '--n_jobs', help='number of jobs running parallel', type=int, default=None)
args = parser.parse_args()
conf = LumosConf()
dump_pth = conf.get('dataset', 'dump_pth_ordinal_with_truc_v1')
# dataloader = DataLoaderOrdinal()
dataloader = DataLoaderOrdinal(dump_pth=dump_pth)
dataloader.load_data_by_interval(interval=1)
rank_data = dataloader.get_data_rankize()
# dataset options
op_truncate = [True, False]
op_ordinal = [True, False]
# model options
op_max_depth = [3, 4, 5]
op_n_estimators = [10, 40, 70, 100]
op_criterion = ['mse', 'mae']
op_max_features = ['auto', 'sqrt', 'log2', 0.5]
conf = LumosConf()
redundant_feat_idxes = []
for idx in range(len(avg_cof)):
if idx not in valid_idxes: continue
if idx in redundant_feat_idxes: continue
selected_feat_idxes.append(idx)
for idx_2 in range(len(avg_cof[idx])):
if idx_2 in selected_feat_idxes: continue
if idx_2 in redundant_feat_idxes: continue
if avg_cof[idx][idx_2] > cof_threshold and idx_2 != idx:
redundant_feat_idxes.append(idx_2)
return selected_feat_idxes
if __name__ == "__main__":
from conf import LumosConf
conf = LumosConf()
dump_pth = conf.get('dataset', 'dump_pth_ordinal_with_truc_v1')
dataloader = DataLoaderOrdinal(dump_pth=dump_pth)
dataloader.load_data_by_interval(interval=1)
data = dataloader.get_data()
to_select_scale = 'large'
metrics_data = []
for wl, wl_data in data['1'].items():
scale_data = wl_data[to_select_scale]
# metrics_data.append(random.sample(scale_data, 1)[0].metrics)
metrics_data.append(scale_data[1].metrics)
feature_idxes = select_features(metrics_data)
# ana_metrics(metrics_data)
print('%d features selected: %s' % (len(feature_idxes), feature_idxes))
import os
import sys
import json
import dill
import random
import pickle
import numpy as np
from utils import *
from conf import LumosConf
from collections import defaultdict
from data_loader_ordinal import DataLoaderOrdinal
from third_party.keras_lstm_vae.lstm_vae import create_lstm_vae
if __name__ == "__main__":
conf = LumosConf()
dump_pth = conf.get('dataset', 'dump_pth_ordinal')
dataloader = DataLoaderOrdinal(dump_pth=dump_pth)
dataloader.load_data()
norm_metrics = normalize_metrics(metrics)
self.__data[workload][vendor].append(
RecordEntry(inst_type, scale, norm_metrics, jct, ts))
def __load_data_from_file(self):
with open(self.dump_pth, 'rb') as fd:
self.__data = dill.load(fd)
def get_data(self):
return self.__data
if __name__ == "__main__":
conf = LumosConf()
dump_pth = None
if conf.get('dataset', 'no_dup'):
dump_pth = conf.get('dataset', 'dump_pth_no_dup')
else:
dump_pth = conf.get('dataset', 'dump_pth')
#data_loader = DataLoader(dump_pth=dump_pth)
data_loader = DataLoader(no_dup=True)
data_loader.load_data()
data = data_loader.get_data()
print(len(data))
print(data.keys())
print(len(data['hadoop_aggregation']['alibaba']))
print(len(data['hadoop_aggregation']['huawei']))
print(len(data['hadoop_aggregation']['tencent']))
# print(len(data['hadoop_aggregation']['ucloud']))
with open(dump_pth, 'wb') as fd:
dill.dump(data, fd)
def get_train_test_data(self,
train_scale='tiny',
test_wl='',
flag='single'):
'''
get the training data that profiled on a concrete instance type
param:
@t_inst_type: the instance type that is used for profiling
@test_wl: the workload that is to be used for testing
'''
rankize_data = self.get_data_rankize()
assert test_wl in self.__data['1'] or test_wl in (
'HiBench', 'BigBench'), 'invalid test workload'
assert flag in ('single',
'multi'), 'indicating single/multi testing workloads'
def is_test_wl(wl):
if flag == 'single':
return wl == test_wl
else:
if test_wl == 'BigBench':
return 'hive' in wl
elif test_wl == 'HiBench':
return 'hive' not in wl
conf = LumosConf()
truncate = conf.get('dataset', 'truncate')
fft_stat_encoder = FFTStatEncoder(truncate=truncate)
train_data = defaultdict(lambda: defaultdict(lambda: {
'X': [],
'Y': []
}))
test_data = defaultdict(lambda: defaultdict(lambda: \
defaultdict(lambda: defaultdict(lambda: {
'X': [],
'Y': []
}))))
predict_scales = ['tiny', 'small', 'large', 'huge']
if train_scale == 'small':
predict_scales.remove('tiny')
for rnd, rnd_data in rankize_data.items():
for wl, wl_data in rnd_data.items():
if is_test_wl(wl): continue
for record1 in wl_data[train_scale]:
t_inst_type = record1.inst_type
test_conf = conf.get_inst_detailed_conf(t_inst_type,
format='list')
test_metrics_vec = fft_stat_encoder.encode(
record1.metrics,
record1.raw_metrics,
sampling_interval=self.sampling_interval)
for scale in predict_scales:
target_scale = conf.get_scale_id(scale)
for record2 in wl_data[scale]:
target_conf = conf.get_inst_detailed_conf(
record2.inst_type, format='list')
target_rank = record2.rank
target_jct = record2.jct
X = test_conf.copy()
X.extend(target_conf)
X.append(target_scale)
X.extend(test_metrics_vec)
train_data[rnd][t_inst_type]['X'].append(X)
if self.ordinal:
train_data[rnd][t_inst_type]['Y'].append(
target_rank)
else:
train_data[rnd][t_inst_type]['Y'].append(
target_jct)
for rnd, rnd_data in rankize_data.items():
for wl, wl_data in rnd_data.items():
if not is_test_wl(wl): continue
# wl_data = rnd_data[test_wl]
for record1 in wl_data[train_scale]:
t_inst_type = record1.inst_type
test_conf = conf.get_inst_detailed_conf(t_inst_type,
format='list')
test_metrics_vec = fft_stat_encoder.encode(
record1.metrics,
record1.raw_metrics,
sampling_interval=self.sampling_interval)
for scale in predict_scales:
target_scale = conf.get_scale_id(scale)
for record2 in wl_data[scale]:
target_conf = conf.get_inst_detailed_conf(
record2.inst_type, format='list')
target_rank = record2.rank
target_jct = record2.jct
X = test_conf.copy()
X.extend(target_conf)
X.append(target_scale)
X.extend(test_metrics_vec)
test_data[wl][rnd][t_inst_type][scale]['X'].append(
X)
if self.ordinal:
test_data[wl][rnd][t_inst_type][scale][
'Y'].append(target_rank)
else:
test_data[wl][rnd][t_inst_type][scale][
'Y'].append(target_jct)
return train_data, test_data