def mean_df():
def avg_ondemand(onds, arg):
ond = dvfsModel()
ond.loadData(filename='data/dvfs/ondemand/' + onds[0],
arg_num=arg,
method='constTime')
df = ond.dataFrame
for f in onds[1:]:
ond.loadData(filename='data/dvfs/ondemand/' + onds[0],
arg_num=arg,
method='constTime')
df['energy'] += ond.dataFrame['energy']
df['energy'] /= len(onds)
return df
ondemand = avg_ondemand(['ferret_completo_2.pkl', 'ferret_completo_3.pkl'],
6)
pw_model = powerModel('data/ipmi_2-32_cpuload.pw')
perf_model = performanceModel('data/dataframes/' + 'completo_ferret_3.pkl',
'data/svr/' + 'completo_ferret_3.pkl')
en_model = energyModel(pw_model, perf_model)
ond = ondemand[['in', 'thr', 'time', 'energy']].sort_values(['in', 'thr'])
nthreads = ond['thr'].unique().shape[0]
ond = pd.crosstab(ond['in'], ond['thr'], ond['energy'], aggfunc=min)
df = en_model.realMinimalEnergy().sort_values('in_cat')['energy']
df = pd.concat([df] * nthreads, axis=1)
ond = pd.DataFrame(ond.values / df.values, columns=ond.columns)
ond.plot.bar()
plt.plot([-1, 6], [1, 1], '--', color='k', label='proposto')
plt.title('Ferret')
plt.tight_layout()
plt.savefig('fotos/comp2/ferret.png')
def createPerformanceModels(appname=None):
for p, idx in zip(parsecapps, parsecapps_argnum):
if appname and not appname in p:
continue
perf_model = performanceModel()
df = perf_model.loadData(filename='data/performance_model/' + p,
arg_num=idx,
verbose=0,
method='constTime')
if 'fluid' in p:
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['thr'].isin([1, 2, 4, 8, 16, 32])]
if 'x264' in p:
perf_model.dataFrame['in_cat'] = 6 - perf_model.dataFrame['in_cat']
if len(df['in_cat']) > 5: #limit to 5 inputs
cats = perf_model.dataFrame['in_cat'].unique()[-5:]
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['in_cat'].isin(cats)]
perf_model.fit(C_=10e3, gamma_=0.1)
scores = perf_model.crossValidate(method='mpe')
perf_model.saveDataframe('data/dataframes/' + p)
perf_model.saveSVR('data/svr/' + p)
print("Program", p)
print(df.head(5))
print("MPE ", perf_model.error(method='mpe') * 100)
print("MAE ", perf_model.error(method='mae'))
print("CrossValidation ", np.mean(scores) * 100, scores)
def figures(appname=None, energy=True, in_cmp=3):
from energyOptimal import plotData
for app, title in zip(parsec_models, titles):
if (appname and not appname in app) or (not app):
continue
pw_model = powerModel('data/ipmi_2-32_cpuload.pw')
perf_model = performanceModel('data/dataframes/' + app,
'data/svr/' + app)
en_model = energyModel(pw_model,
perf_model,
freq_range_=np.arange(1.2e6, 2.3e6, 0.1e6) /
1e6)
plotData.setProps(xlabel='Frequencies (GHz)',
ylabel='Active threads',
zlabel='Energy (kJ)' if energy else 'Time (s)',
title=title)
df_ = perf_model.dataFrame[perf_model.dataFrame['in_cat'] ==
in_cmp].sort_values(['freq', 'thr'])
df_pred_ = en_model.dataFrame[en_model.dataFrame['in_cat'] ==
in_cmp].sort_values(['freq', 'thr'])
# df_pred_= df_pred_[df_pred_['thr'].isin(list(range(8,33,2)))]
# df_= df_[df_['thr'].isin(list(range(8,33,2)))]
plotData.plot3D(x=df_['freq'].unique(),
y=df_['thr'].unique(),
z=df_['energy'].values /
1e3 if energy else df_['time'].values,
points=True,
legend='Measurements')
plotData.plot3D(x=df_pred_['freq'].unique(),
y=df_pred_['thr'].unique(),
z=df_pred_['energy_model'].values /
1e3 if energy else df_pred_['time'].values,
points=False,
legend='Model')
plotData.ax.view_init(30, 60)
if 'HPL' in app:
plotData.ax.set_zlim(0, 15)
aux = 'energy' if energy else 'time'
plotData.savePlot('fotos/{}/{}.png'.format(aux, app), showLegend=True)
def createPerformanceModel(path, idx, save_df, save_svr):
perf_model = performanceModel()
perf_model.loadData(filename=path,
arg_num=int(idx),
verbose=0,
method='constTime',
freqs_filter=get_range(args.freqs),
thrs_filter=get_range(args.thrs))
if args.ins:
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['in_cat'].isin(get_range(args.ins))]
print("Program", path)
print(perf_model.dataFrame.head(5))
perf_model.fit(C_=10e3, gamma_=0.1)
print("MPE ", perf_model.error(method='mpe') * 100)
print("MAE ", perf_model.error(method='mae'))
scores = perf_model.crossValidate(method='mpe')
print("CrossValidation ", np.mean(scores) * 100, scores)
perf_model.saveDataframe(save_df)
perf_model.saveSVR(save_svr)
def visualizePeformanceModel(path_df, path_svr, title_):
perf_model = performanceModel(path_df, path_svr)
df = perf_model.dataFrame.sort_values(['freq', 'thr'])
freqs = get_range(args.freqs) if args.freqs else df['freq'].unique()
thrs = get_range(args.thrs) if args.thrs else df['thr'].unique()
ins = get_range(args.ins) if args.ins else df['in_cat'].unique()
df_pred = perf_model.estimate_(freqs, thrs, ins,
dataframe=True).sort_values(['freq', 'thr'])
print(df_pred.head(5))
def update_data(val):
d = int(val)
plotData.setProps(xlabel='Frequencies (GHz)',
ylabel='Active threads',
zlabel='Time (S)',
title=title_)
df_ = df[df['in_cat'] == d]
df_pred_ = df_pred[df_pred['in_cat'] == d]
df_ = df_[df_['thr'].isin(df_pred_['thr'].unique())]
if not df_.empty:
plotData.plot3D(x=df_['freq'].unique(),
y=df_['thr'].unique(),
z=df_['time'].values)
plotData.plot3D(x=df_pred_['freq'].unique(),
y=df_pred_['thr'].unique(),
z=df_pred_['time'].values,
points=False)
plotData.update_user = update_data
plotData.createSlider(label_='in',
valmin_=df_pred['in_cat'].min(),
valmax_=df_pred['in_cat'].max())
plotData.ax.view_init(30, 60)
plotData.plotShow()
def createReducedPerformanceModel(path,
arg_num,
title_='',
save_df='',
save_svr=''):
perf_model = performanceModel()
perf_model.loadData(filename=path, arg_num=int(arg_num))
cats = perf_model.dataFrame['in_cat'].unique()[-5:]
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['in_cat'].isin(cats)]
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['freq'] != 2.3]
if 'fluid' in path:
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['thr'].isin([1, 2, 4, 8, 16, 32])]
if 'x264' in path:
perf_model.dataFrame['in_cat'] = 6 - perf_model.dataFrame['in_cat']
df_ori = perf_model.dataFrame.sort_values(['freq', 'thr', 'in_cat']).copy()
x = []
y_time = []
y_en = []
less_5 = 0
for train_sz in range(0, perf_model.dataFrame.shape[0], 100)[1:]:
# print("Program", path)
aux = perf_model.fit(C_=10e3,
gamma_=0.1,
train_size_=train_sz,
dataframe=True)
aux = pd.merge(aux[['freq', 'thr', 'in_cat']], df_ori)
perf_model.estimate(df_ori[['freq', 'thr', 'in_cat']],
dataframe=True).sort_values(
['freq', 'thr', 'in_cat'])
x.append(train_sz)
y_time.append(perf_model.error() * 100)
y_en.append(aux['energy'].sum() / 1e6)
# print(y_en[-1])
# print( x[-1], y_time[-1] )
if y_time[-1] <= 6 and less_5 == 0:
less_5 = y_time[-1]
print('%s_%i.pkl' % (title_, train_sz))
perf_model.saveDataframe('data/dataframes/%s_%i.pkl' %
(title_, train_sz))
perf_model.saveSVR('data/svr/%s_%i.pkl' % (title_, train_sz))
break
# scores= perf_model.crossValidate(method='mpe')
# print("CrossValidation ", np.mean(scores)*100, scores)
fig, ax1 = plt.subplots()
ax1.plot(x, y_time)
# ax1.plot([min(x),max(x)],[less_5, less_5],'-')
ax1.set_ylabel('Mean error (%)')
ax2 = ax1.twinx()
ax2.plot(x, y_en)
ax2.set_ylabel('Energy (KJ)')
plt.xlabel('Train size')
plt.title(title_)
plt.savefig('fotos/over/%s.png' % title_)
def comparation(appname=None,
proposed_bar=False,
relative=True,
thrs_filter=[]):
row = []
for title, dvfs, model, arg in zip(titles, parsec_dvfs, parsec_models,
parsecapps_argnum):
if 'freq' in model or not model:
continue
if appname and not appname in dvfs:
continue
ondemand = dvfsModel()
ondemand.loadData(filename='data/dvfs/ondemand/' + dvfs,
arg_num=arg,
method='constTime')
pw_model = powerModel('data/ipmi_2-32_cpuload.pw')
perf_model = performanceModel('data/dataframes/' + model,
'data/svr/' + model)
en_model = energyModel(pw_model, perf_model)
#TODO verify if arguments match
ond = ondemand.dataFrame[['in', 'thr', 'time', 'energy']]
ond = pd.merge(ond, perf_model.dataFrame[[
'in', 'in_cat'
]]).drop_duplicates().sort_values(['in_cat', 'thr'])
if thrs_filter:
ond = ond[ond['thr'].isin(thrs_filter)]
ond_en = pd.crosstab(ond['in_cat'],
ond['thr'],
ond['energy'],
aggfunc=min)
ond_time = pd.crosstab(ond['in_cat'],
ond['thr'],
ond['time'],
aggfunc=min)
min_df = en_model.realMinimalEnergy().sort_values('in_cat')
if proposed_bar:
ond_en['proposed'] = min_df['energy'].values
ond_time['proposed'] = min_df['time'].values
if relative:
aux = pd.concat([min_df['energy']] * (len(ond_en.columns)), axis=1)
saving = pd.DataFrame(ond_en.values - aux.values,
columns=ond_en.columns)
ond_en = pd.DataFrame(ond_en.values / aux.values,
columns=ond_en.columns)
aux = pd.concat([min_df['time']] * (len(ond_en.columns)), axis=1)
ond_time = pd.DataFrame(ond_time.values / aux.values,
columns=ond_time.columns)
row.append([
dvfs,
ond_en.max(axis=1).mean(),
ond_en.mean(axis=1).mean(), ond_en[32].mean(),
ond_en.min(axis=1).mean(),
ond_time.max(axis=1).mean(),
ond_time.mean(axis=1).mean(), ond_time[32].mean(),
ond_time.min(axis=1).mean(),
saving.max(axis=1).sum(),
saving.median(axis=1).sum(), saving[32].sum(),
saving.min(axis=1).sum(),
perf_model.dataFrame.energy.sum()
])
ond_en.plot.bar(figsize=(12, 7))
if relative:
plt.plot([-1, 6], [1, 1], '--', color='k', label='proposed')
plt.plot([-1, 6], [ond_en.mean().mean(),
ond_en.mean().mean()],
':',
color='k',
label='average gain')
plt.xlabel('Inputs', fontsize=18)
plt.ylabel('Energy relative' if relative else 'Eenergy (J)',
fontsize=18)
plt.xticks(fontsize=18)
plt.yticks(fontsize=18)
plt.title(title, fontsize=22)
plt.legend(loc='center right', bbox_to_anchor=(1.2, 0.55), fontsize=12)
plt.tight_layout()
plt.savefig('fotos/relative/%s.png' % dvfs)
# plt.show()
if relative and not proposed_bar:
df = pd.DataFrame(row,
columns=[
'app', 'max_en', 'mean_en', '32_en', 'min_en',
'max_time', 'mean_time', '32_time', 'min_time',
'max_save', 'median_save', '32_save', 'min_save',
'train_energy'
])
df[[
'max_en', 'mean_en', '32_en', 'min_en', 'max_time', 'mean_time',
'32_time', 'min_time'
]] -= 1
df[[
'max_en', 'mean_en', '32_en', 'min_en', 'max_time', 'mean_time',
'32_time', 'min_time'
]] *= 100
df['max_run'] = df['train_energy'] / df['max_save']
df['median_run'] = df['train_energy'] / df['median_save']
df['32_run'] = df['train_energy'] / df['32_save']
df['min_run'] = df['train_energy'] / df['min_save']
df = df.sort_values('32_en', ascending=False)
df = pd.concat((df,
pd.DataFrame([['mean'] + list(df.mean().values)],
columns=df.columns)))
df.to_csv('tables/relative.csv')
print(df)
def createReducedPerformanceModel2(path,
arg_num,
title_='',
save_df='',
save_svr=''):
perf_model = performanceModel()
perf_model.loadData(filename=path, arg_num=int(arg_num))
cats = perf_model.dataFrame['in_cat'].unique()[-5:]
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['in_cat'].isin(cats)]
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['freq'] < 2.3]
if 'fluid' in path:
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['thr'].isin([1, 2, 4, 8, 16, 32])]
if 'x264' in path:
perf_model.dataFrame['in_cat'] = 6 - perf_model.dataFrame['in_cat']
df_ori = perf_model.dataFrame.sort_values(['freq', 'thr', 'in_cat']).copy()
ori = perf_model.dataFrame.copy()
freqs = perf_model.dataFrame['freq'].unique()
x = []
y_time = []
y_en = []
for f in range(1, len(freqs), 1):
use_freq = list(freqs[:int(f)]) + list(freqs[-int(f):])
perf_model.dataFrame = perf_model.dataFrame[
perf_model.dataFrame['freq'].isin(use_freq)]
# print(perf_model.dataFrame['freq'].unique())
aux = perf_model.fit(C_=10e3,
gamma_=0.1,
train_size_=0.9,
dataframe=True)
aux = pd.merge(aux[['freq', 'thr', 'in_cat']], df_ori)
perf_model.dataFrame = ori.copy()
df_est = perf_model.estimate(df_ori[['freq', 'thr', 'in_cat']],
dataframe=True).sort_values(
['freq', 'thr', 'in_cat'])
error = sum((abs(df_est['time'] - df_ori['time'])) /
df_ori['time']) / df_ori.shape[0] * 100
x.append(aux.shape[0])
y_time.append(error)
y_en.append(aux['energy'].sum() / 1e6)
# scores= perf_model.crossValidate(method='mpe')
print('%s_%i.pkl' % (title_, f), aux.shape, aux['energy'].sum() / 1e6,
error,
perf_model.error() * 100)
print(use_freq)
perf_model.saveDataframe('data/dataframes/%s_%i.pkl' % (title_, f))
perf_model.saveSVR('data/svr/%s_%i.pkl' % (title_, f))
fig, ax1 = plt.subplots()
ax1.plot(x, y_time)
ax1.set_ylabel('Mean error (%)')
ax2 = ax1.twinx()
ax2.plot(x, y_en)
ax2.set_ylabel('Energy (KJ)')
plt.xlabel('Train size')
plt.title(title_)
plt.savefig('fotos/over/%s.png' % title_)
sr = r.split(',')
if len(sr) >= 2:
start = float(sr[0])
stop = float(sr[1])
if len(sr) == 3:
step = float(sr[2])
else:
step = 1.0
_range += list(np.arange(start, stop, step))
elif len(sr) > 0:
_range += list([float(sr[0])])
return _range
pw_model = powerModel(args.path_power)
perf_model = performanceModel(args.path_dataframe, args.path_svr)
en_model = energyModel(pw_model,
perf_model,
thr_range_=get_range(args.thrs),
freq_range_=get_range(args.freqs))
df = perf_model.dataFrame.sort_values(['freq', 'thr'])
df_pred = en_model.dataFrame.sort_values(['freq', 'thr'])
print(en_model.minimalEnergy())
def update_data(val):
d = int(val)
plotData.setProps(xlabel='Frequencies (GHz)',
ylabel='Active threads',
zlabel='Energy (KJ)',