def run():
"""
:param app_name: HPC application
:param perf_coln: performance name to be optimized
:param num_core: number of CPU cores
:param num_node: number of computing nodes
:param rand_seed: random seed
:param num_smpl: number of samples
:param pool_size: pool size
:param num_iter: number of iterations
:param prec_rand: precentage of random samples
"""
try:
cm.init()
app_name = cm.app_name
perf_coln = cm.perf_coln
num_smpl = cm.num_smpl
pool_size = cm.pool_size
num_iter = cm.num_iter
prec_rand = cm.prec_rand
if (app_name == "lv"):
conf_colns = data.lv_conf_colns
elif (app_name == "hs"):
conf_colns = data.hs_conf_colns
num_rand = int(num_smpl * prec_rand)
nspi = int((num_smpl - num_rand) / num_iter)
# pool_df = data.gen_smpl(app_name, pool_size)
# conf_df = pool_df.head(num_rand)
conf_df = data.gen_smpl(app_name, num_rand)
train_df = cm.measure_perf(conf_df)
for iter_idx in range(num_iter):
num_curr = num_smpl - nspi * (num_iter - 1 - iter_idx)
pool_df = data.gen_smpl(app_name, pool_size)
pred_top_smpl = learn.whl_pred_top_eval(train_df, pool_df, conf_colns, perf_coln, num_smpl, 0)
pred_top_smpl = pred_top_smpl.sort_values([perf_coln]).reset_index(drop=True)
new_conf_df = pred_top_smpl[conf_colns].head(nspi)
conf_df = tool.df_union(conf_df, new_conf_df)
last = nspi
while (conf_df.shape[0] < num_curr):
last = last + 1
new_conf_df = pred_top_smpl[conf_colns].head(last)
conf_df = tool.df_union(conf_df, new_conf_df)
new_train_df = cm.measure_perf(new_conf_df)
train_df = tool.df_union(train_df, new_train_df)
data.df2csv(train_df, app_name + "_train.csv")
mdl_chk, mdl = learn.train_mdl_chk(train_df, conf_colns, perf_coln)
top_df = cm.find_top('ALe', (mdl_chk, mdl, ), conf_colns, perf_coln, train_df)
cm.test(train_df, conf_colns, perf_coln)
cm.finish(train_df, top_df)
except:
traceback.print_exc()
def run():
"""
:param app_name: HPC application
:param perf_coln: performance name to be optimized
:param num_core: number of CPU cores
:param num_node: number of computing nodes
:param rand_seed: random seed
:param num_smpl: number of samples
"""
try:
cm.init()
app_name = cm.app_name
perf_coln = cm.perf_coln
num_smpl = cm.num_smpl
if (app_name == "lv"):
conf_colns = data.lv_conf_colns
in_conf_colns = data.lv_in_conf_colns
in_params = data.lmp_in_params
elif (app_name == "hs"):
conf_colns = data.hs_conf_colns
in_conf_colns = data.hs_in_conf_colns
in_params = data.ht_in_params
train_in_fn = cm.app_in_name(app_name) + "_time.csv"
train_in_df = data.csv2df(train_in_fn, in_conf_colns)
train_in_df = data.get_exec_mach_df(
data.get_runnable_df(train_in_df, in_conf_colns))
in_mdl_chk, in_mdl = learn.train_mdl_chk(train_in_df, in_conf_colns,
perf_coln)
conf_df = data.gen_smpl(app_name, num_smpl)
train_df = cm.measure_perf(conf_df)
data.df2csv(train_df, app_name + "_train.csv")
train_intrmdt_df = learn.add_layer_shrt_pred(train_df, in_params, conf_colns, perf_coln, \
in_mdl_chk, in_mdl)
mdl_chk, mdl = learn.train_mdl_chk(train_intrmdt_df,
conf_colns + [perf_coln + '0'],
perf_coln)
top_df = cm.find_top('cmtl_wh', (
in_mdl_chk,
in_mdl,
mdl_chk,
mdl,
), conf_colns, perf_coln, train_df)
test_df = data.csv2df(app_name + "_time.csv", conf_colns)
test_df = data.get_exec_mach_df(
data.get_runnable_df(test_df, conf_colns))
test_intrmdt_df = learn.add_layer_shrt_pred(test_df, in_params, conf_colns, perf_coln, \
in_mdl_chk, in_mdl)
'''
pred_top, err, rs = learn.whl_pred_top_eval(train_intrmdt_df, test_intrmdt_df, \
conf_colns + [perf_coln+'0'], perf_coln, 10)
data.df2csv(pred_top, app_name + "_test.csv")
data.df2csv(rs, app_name + "_rs")
data.df2csv(err, app_name + "_err.csv")
'''
cm.finish(train_df, top_df)
except:
traceback.print_exc()
def run():
"""
:param app_name: HPC application
:param perf_coln: performance name to be optimized
:param num_core: number of CPU cores
:param num_node: number of computing nodes
:param rand_seed: random seed
:param num_smpl: number of samples
:param pool_size: pool size
:param num_iter: number of iterations
:param prec_rand: precentage of random samples
:param prec_init: precentage of initial samples replaced by equivalent samples
"""
try:
cm.init()
app_name = cm.app_name
perf_coln = cm.perf_coln
num_smpl = cm.num_smpl
pool_size = cm.pool_size
num_iter = cm.num_iter
prec_rand = cm.prec_rand
prec_init = cm.prec_init
if (app_name == "lv"):
conf_colns = data.lv_conf_colns
conf1_colns = data.lmp_conf_colns
conf2_colns = data.vr_conf_colns
elif (app_name == "hs"):
conf_colns = data.hs_conf_colns
conf1_colns = data.ht_conf_colns
conf2_colns = data.sw_conf_colns
num_rand = int(num_smpl * prec_rand)
# pool_df = data.gen_smpl(app_name, pool_size)
# conf_df = pool_df.head(num_rand)
conf_df = data.gen_smpl(app_name, num_rand)
train_df = cm.measure_perf(conf_df)
num_init = int(num_smpl * prec_init)
pool1_df = data.gen_smpl(cm.app1_name(app_name), num_init * 100)
conf1_df = pool1_df.head(num_init)
train1_df = cm.measure_perf(conf1_df)
pool2_df = data.gen_smpl(cm.app2_name(app_name), num_init * 100)
conf2_df = pool2_df.head(num_init)
train2_df = cm.measure_perf(conf2_df)
avg_mach_time = data.sa_mach_time(train_df) / num_rand
avg_sprt_mach_time = (data.app_mach_time(train1_df) + \
data.app_mach_time(train2_df)) / num_init
factor = max(1, avg_mach_time / avg_sprt_mach_time)
if (factor > 1):
num_sprt = int(num_init * factor)
new_conf1_df = pool1_df.head(num_sprt).tail(num_sprt - num_init)
new_train1_df = cm.measure_perf(new_conf1_df)
train1_df = tool.df_union(train1_df, new_train1_df)
new_conf2_df = pool2_df.head(num_sprt).tail(num_sprt - num_init)
new_train2_df = cm.measure_perf(new_conf2_df)
train2_df = tool.df_union(train2_df, new_train2_df)
pool_df = data.gen_smpl(app_name, pool_size)
pred_top_smpl = learn.sprt_pred_top_eval(train1_df, train2_df, pool_df,
conf1_colns, conf2_colns,
conf_colns, perf_coln,
num_smpl, 0)
nspi = int((num_smpl - num_init - num_rand) / num_iter)
for iter_idx in range(num_iter):
num_curr = num_smpl - num_init - nspi * (num_iter - 1 - iter_idx)
pred_top_smpl = pred_top_smpl.sort_values(
[perf_coln]).reset_index(drop=True)
new_conf_df = pred_top_smpl[conf_colns].head(nspi)
conf_df = tool.df_union(conf_df, new_conf_df)
last = nspi
while (conf_df.shape[0] < num_curr):
last = last + 1
new_conf_df = pred_top_smpl[conf_colns].head(last)
conf_df = tool.df_union(conf_df, new_conf_df)
new_train_df = cm.measure_perf(new_conf_df)
train_df = tool.df_union(train_df, new_train_df)
if (iter_idx < num_iter - 1):
pool_df = data.gen_smpl(app_name, pool_size)
pred_top_smpl = learn.whl_pred_top_eval(
train_df, pool_df, conf_colns, perf_coln, num_smpl, 0)
data.df2csv(train_df, app_name + "_train.csv")
mdl_chk, mdl = learn.train_mdl_chk(train_df, conf_colns, perf_coln)
top_df = cm.find_top('TaLeC', (
mdl_chk,
mdl,
), conf_colns, perf_coln, train_df)
cm.test(train_df, conf_colns, perf_coln)
cm.finish(train_df, top_df)
except:
traceback.print_exc()