def pairOpt(workloads):
nStep = 1000
nTry = 10
dt = 1
# machines = range(1, int(len(workloads)/2) + 1)
idx = lambda hi, ni: ni + hi * len(machines)
ridx = lambda x: (x // len(machines), x % len(machines))
models = {}
# param = {
# 'loss': 'ls',
# 'learning_rate': .1,
# 'alpha': 0.9,
# 'max_depth': 4,
# 'n_estimators': 100,
# 'm': 10000,
# }
perflog = {
'realized': [],
'predicted': [],
}
totrain = []
for hi in machines:
model = model_xgb.XGBoost()
model.init(**params)
totrain.append((model, hi, nTrain, dt, libdata.apps))
lmodels = pool.map(trainModel, totrain)
for i in range(len(lmodels)):
models[i + 1] = lmodels[i]
cache = {}
print(workloads)
stdwl = list(stdWorkloads(workloads))
print(stdwl)
table = []
for i in range(len(stdwl)):
hi = i + 1
app0, app1 = stdwl[i]
p = _pair_predict(models[hi], hi, app0, app1, cache=cache, dt=dt)
table.append({'power': p['power_0'], 'app': app0})
table.append({'power': p['power_1'], 'app': app1})
tableidx = [idx(hi, ni) for hi in machines for ni in range(2)]
table = pd.DataFrame(table, index=tableidx)
while nStep > 0:
nStep -= 1
perflog['realized'].append(
oneRun([
table['app'].loc[idx(hi, ni)] for hi in machines
for ni in range(2)
])['pkgpwr'])
perflog['predicted'].append(table['power'].sum())
tsorted = table.sort_values(by='power', ascending=False)
# print(tsorted)
ok = False
for i, j in [(k, len(tsorted) - l - 1) for k in range(nTry)
for l in range(nTry)]:
h0, n0 = ridx(tsorted.index[i])
h1, n1 = ridx(tsorted.index[j])
if tsorted.loc[idx(h0, n0), 'app'] == tsorted.loc[idx(h1, n1),
'app']:
continue
#print(tsorted.columns.values)
pbefore = tsorted['power'].loc[[
idx(h0, n0),
idx(h0, 1 - n0),
idx(h1, n1),
idx(h1, 1 - n1)
]].sum()
a00, a01 = tsorted.loc[idx(h0, 0), 'app'], tsorted.loc[idx(h0, 1),
'app']
a10, a11 = tsorted.loc[idx(h1, 0), 'app'], tsorted.loc[idx(h1, 1),
'app']
a0before = [a00, a01]
a1before = [a10, a11]
a0after = a0before[:]
a1after = a1before[:]
a0after[n0] = a1before[n1]
a1after[n1] = a0before[n0]
if h0 == h1:
a0after[n1] = a1after[n1]
a1after[n0] = a0after[n0]
# print(a0before, a0after, a1before, a1after)
pa0, pa1 = _pair_predict(models[h0],
h0,
a0after[0],
a0after[1],
cache=cache,
dt=dt)
pb0, pb1 = _pair_predict(models[h1],
h1,
a1after[0],
a1after[1],
cache=cache,
dt=dt)
if pa0 + pa1 + pb0 + pb1 < pbefore:
tsorted.set_value(idx(h0, 0), 'app', a0after[0])
tsorted.set_value(idx(h0, 1), 'app', a0after[1])
tsorted.set_value(idx(h1, 0), 'app', a1after[0])
tsorted.set_value(idx(h1, 1), 'app', a1after[1])
tsorted.set_value(idx(h0, 0), 'power', pa0)
tsorted.set_value(idx(h0, 1), 'power', pa1)
tsorted.set_value(idx(h1, 0), 'power', pb0)
tsorted.set_value(idx(h1, 1), 'power', pb1)
ok = True
break
if not ok:
break
table = tsorted
perflog = pd.DataFrame(perflog)
fig, ax = plt.subplots()
ax.plot(perflog.index, perflog['realized'], 'r')
ax.plot(perflog.index, perflog['predicted'], 'b')
fig.suptitle(','.join(workloads))
#fig.savefig(optpdf, format='pdf')
return [
table['app'].loc[idx(hi, ni)] for hi in machines for ni in range(2)
]
allres = []
pred_time = 0
train_time = 0
for eval_times in range(10):
print("start: %s\t%d" % (str(datetime.datetime.now()), eval_times))
pool = mp.Pool(mp.cpu_count())
totrain = []
#apps_train = rand.sample(libdata.apps, 9)
apps_train = apps_npb
temp = list(set(libdata.apps) - set(list(apps_train)))
apps_validation = temp
print("validation apps: ", apps_validation)
train_start = time.time()
for hi in machines:
if ml_method == 'xgb':
model = model_xgb.XGBoost()
if ml_method == 'lr':
model = model_lr.LR()
if ml_method == 'svr':
model = model_svr.SVM()
if ml_method == 'gp':
model = model_gp.GPR()
if ml_method == 'mlp':
model = model_mlp.MLP()
model.init(**params)
totrain.append((model, hi, nTrain, dt, apps_train))
lmodels = pool.map(trainModel, totrain)
print("finish: %s\t%d" % (str(datetime.datetime.now()), eval_times))
for i in range(len(lmodels)):
models[i + 1] = lmodels[i]
train_time += (time.time() - train_start)