def csv_contour(rho_single: float,
sigma_list: List[float],
utilization: float,
perform_param: PerformParameter,
pure_snc: False = False) -> pd.DataFrame:
agg_list = [0] * len(sigma_list)
for i, sigma in enumerate(sigma_list):
agg_list[i] = single_hop_contour(sigma_single=sigma,
rho_single=rho_single,
utilization=utilization,
perform_param=perform_param,
pure_snc=pure_snc)
results_df = pd.DataFrame({
"aggregation": agg_list,
}, index=sigma_list)
if pure_snc:
filename = "contour_{0}_rho_{1}_utilization_{2}_pure".format(
perform_param.to_name_value(), str(rho_single),
str("%.2f" % utilization))
else:
filename = "contour_{0}_rho_{1}_utilization_{2}".format(
perform_param.to_name_value(), str(rho_single),
str("%.2f" % utilization))
results_df.to_csv(filename + '.csv',
index=True,
quoting=csv.QUOTE_NONNUMERIC)
return results_df
def arrival_list_to_csv(prefix: str,
data_frame_creator: Callable,
list_arr_list: List[List[ArrivalDistribution]],
ser_list: List[ConstantRateServer],
server_index: int,
perform_param: PerformParameter,
opt_method: OptMethod = None,
suffix="_adjusting_arrivals") -> pd.DataFrame:
filename = prefix + perform_param.to_name()
if opt_method is None:
data_frame: pd.DataFrame = data_frame_creator(
list_arr_list=list_arr_list,
ser_list=ser_list,
server_index=server_index,
perform_param=perform_param)
else:
data_frame: pd.DataFrame = data_frame_creator(
list_arr_list=list_arr_list,
ser_list=ser_list,
server_index=server_index,
opt_method=opt_method,
perform_param=perform_param)
filename += "_" + list_arr_list[0][0].to_name()
data_frame.to_csv(filename + suffix + '.csv',
index=True,
quoting=csv.QUOTE_NONNUMERIC)
return data_frame
def helper_function(delay: float) -> float:
perform_delay = PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=delay)
current_delay_prob = pmoo_explicit(foi=foi,
cross_flows=cross_flows,
ser_list=ser_list,
theta=theta,
perform_param=perform_delay,
indep=indep)
return target_delay_prob - current_delay_prob
def helper_function(delay: float) -> float:
perform_delay = PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=delay)
current_delay_prob = sfa_explicit(
foi=foi,
leftover_service_list=leftover_service_list,
theta=theta,
perform_param=perform_delay,
indep=indep)
return target_delay_prob - current_delay_prob
def new_delay(theta: float, delay: int, a_1: ArrivalDistribution,
a_2: ArrivalDistribution, s_1: ConstantRate,
s_2: TokenBucketConstant, s_3: ConstantRate) -> float:
s_net: Service = Convolve(ser1=Leftover(arr=a_2, ser=s_1),
ser2=Leftover(arr=s_2, ser=s_3),
indep=True)
return evaluate_single_hop(foi=a_1,
s_net=s_net,
theta=theta,
perform_param=PerformParameter(
perform_metric=PerformEnum.DELAY_PROB,
value=delay))
def grid_param_simple_exp(delay: int, opt_method: OptMethod, metric: str,
lamb1_range, lamb2_range, rate1_range,
rate2_range) -> dict:
"""Choose parameters along a grid."""
total_iterations = len(lamb1_range) * len(lamb2_range) * len(
rate1_range) * len(rate2_range)
param_array = np.empty([total_iterations, 4])
res_array = np.empty([total_iterations, 2])
i = 0
for lamb1 in lamb1_range:
for lamb2 in lamb2_range:
for rate1 in rate1_range:
for rate2 in rate2_range:
delay_prob = PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=delay)
setting = FatCrossPerform(
arr_list=[DM1(lamb=lamb1),
DM1(lamb=lamb2)],
ser_list=[
ConstantRate(rate=rate1),
ConstantRate(rate=rate2)
],
perform_param=delay_prob)
param_array[i, 0] = lamb1
param_array[i, 1] = lamb2
param_array[i, 2] = rate1
param_array[i, 3] = rate2
# bound, new_bound
res_array[i, 0], res_array[i, 1] = compute_improvement(
setting=setting, opt_method=opt_method, number_l=1)
# This might be a very dangerous condition
if (res_array[i, 1] >= 1 or res_array[i, 0] == nan
or res_array[i, 1] == nan):
res_array[i, ] = nan
if i % floor(total_iterations / 10) == 0:
print("iteration {0} of {1}".format(
i, total_iterations))
i += 1
return two_col_array_to_results(arrival_enum=ArrivalEnum.DM1,
metric=metric,
param_array=param_array,
res_array=res_array,
number_servers=2)
def helper_function(delay: float) -> float:
perform_delay = PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=delay)
current_delay_prob = sfa_tandem_bound(
foi=foi,
leftover_service_list=leftover_service_list,
theta=theta,
perform_param=perform_delay,
p_list=p_list,
e2e_enum=e2e_enum,
indep=indep,
geom_series=True)
return target_delay_prob - current_delay_prob
def helper_function(delay: float) -> float:
perform_delay = PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=delay)
current_delay_prob = pmoo_tandem_bound(
foi=foi,
cross_flows_on_foi_path=cross_flows_on_foi_path,
ser_on_foi_path=ser_on_foi_path,
theta=theta,
perform_param=perform_delay,
e2e_enum=e2e_enum,
indep=indep,
geom_series=True)
return target_delay_prob - current_delay_prob
def helper_function(rate: float):
if opt_method == OptMethod.GRID_SEARCH:
if indep:
single_server = SingleServerBandwidth(
arr_list=arr_list,
s_e2e=ConstantRateServer(rate=rate),
perform_param=PerformParameter(
perform_metric=PerformEnum.DELAY_PROB,
value=target_delay),
indep=True,
geom_series=geom_series)
current_delay_prob = Optimize(setting=single_server,
number_param=1).grid_search(
grid_bounds=[(0.1, 5.0)],
delta=0.1)
else:
single_server = SingleServerBandwidth(
arr_list=arr_list,
s_e2e=ConstantRateServer(rate=rate),
perform_param=PerformParameter(
perform_metric=PerformEnum.DELAY_PROB,
value=target_delay),
indep=False,
geom_series=geom_series)
current_delay_prob = Optimize(setting=single_server,
number_param=2).grid_search(
grid_bounds=[(0.1, 5.0),
(1.1, 5.0)],
delta=0.1)
else:
raise NotImplementedError("This optimization method is not "
"implemented")
return current_delay_prob - target_delay_prob
def standard_delay(theta: float, p: float, delay: int,
a_1: ArrivalDistribution, a_2: ArrivalDistribution,
a_3: ArrivalDistribution, s_1: ConstantRate,
s_2: ConstantRate, s_3: ConstantRate) -> float:
f_3_output: Arrival = Deconvolve(arr=a_3,
ser=Leftover(arr=Deconvolve(arr=a_2,
ser=s_1),
ser=s_2))
s_net: Service = Convolve(ser1=Leftover(arr=a_2, ser=s_1),
ser2=Leftover(arr=f_3_output, ser=s_3),
indep=False,
p=p)
return evaluate_single_hop(foi=a_1,
s_net=s_net,
theta=theta,
perform_param=PerformParameter(
perform_metric=PerformEnum.DELAY_PROB,
value=delay))
def compare_aggregation(aggregations: List[int], sigma_single: float,
rho_single: float, service_rate: float,
perform_param: PerformParameter,
opt_method: OptMethod) -> pd.DataFrame:
dnc_fifo_single = [0.0] * len(aggregations)
const_opt = [0.0] * len(aggregations)
leaky_mass_1 = [0.0] * len(aggregations)
leaky_mass_2_opt = [0.0] * len(aggregations)
exact_mass_2_opt = [0.0] * len(aggregations)
for i, agg in enumerate(aggregations):
dnc_fifo_single[i], const_opt[i], leaky_mass_1[i], leaky_mass_2_opt[
i], exact_mass_2_opt[i] = single_hop_comparison(
aggregation=agg,
sigma_single=sigma_single,
rho_single=rho_single,
service_rate=service_rate * agg,
perform_param=perform_param,
opt_method=opt_method)
results_df = pd.DataFrame(
{
"DNCBound": dnc_fifo_single,
"constBound": const_opt,
"leakyMassOne": leaky_mass_1,
"leakyMassTwo": leaky_mass_2_opt,
"exactMassTwo": exact_mass_2_opt
},
index=aggregations)
filename = (
"regulated_single_{0}_sigma_{1}_rho_{2}_utilization_{3}_{4}").format(
perform_param.to_name_value(), str(sigma_single), str(rho_single),
str("%.2f" % (rho_single / service_rate)), opt_method.name)
results_df.to_csv(filename + '.csv',
index=True,
quoting=csv.QUOTE_NONNUMERIC)
return results_df
return inf
else:
try:
return self.setting_bound.bound(param_list=param_list)
except (ParameterOutOfBounds, OverflowError):
return inf
if __name__ == '__main__':
from fat_tree.fat_cross_perform import FatCrossPerform
from utils.perform_parameter import PerformParameter
from nc_operations.perform_enum import PerformEnum
from nc_service.constant_rate_server import ConstantRate
from nc_arrivals.markov_modulated import MMOOFluid
DELAY_4 = PerformParameter(perform_metric=PerformEnum.DELAY, value=0.0001)
MMOO_1 = MMOOFluid(mu=1.0, lamb=2.2, burst=3.4)
MMOO_2 = MMOOFluid(mu=3.6, lamb=1.6, burst=0.4)
CONST_RATE_1 = ConstantRate(rate=2.0)
CONST_RATE_2 = ConstantRate(rate=0.3)
SIMPLEX_START = np.array([[0.1], [0.3]])
# SIMPLEX_START = np.array([[100], [200]])
SIMPLEX_START_NEW = np.array([[0.1, 2.0], [0.3, 1.2], [0.4, 1.1]])
SIMPLEX_RAND = InitialSimplex(parameters_to_optimize=1).uniform_dist(
max_theta=0.6, max_l=2.0)
NM_PARAM_SET = NelderMeadParameters()
SETTING = FatCrossPerform(
"""Computed some examples from the project"""
from fat_tree.fat_cross_perform import FatCrossPerform
from nc_arrivals.qt import DM1
from nc_operations.perform_enum import PerformEnum
from nc_service.constant_rate_server import ConstantRate
from optimization.optimize import Optimize
from utils.perform_parameter import PerformParameter
if __name__ == '__main__':
PROB_VALUES = [0.5, 0.4, 0.3, 0.2, 0.1, 0.05, 0.01]
for p in PROB_VALUES:
DELAY_TIME = PerformParameter(perform_metric=PerformEnum.DELAY,
value=p)
EXAMPLE = FatCrossPerform(arr_list=[DM1(lamb=1.0)],
ser_list=[ConstantRate(rate=2.0)],
perform_param=DELAY_TIME)
print(
Optimize(setting=EXAMPLE, print_x=False,
show_warn=True).grid_search(bound_list=[(0.01, 1.1)],
delta=0.01))
def csv_single_param_exp_lower(start_time: int,
perform_param: PerformParameter,
mc_dist: MonteCarloDist,
sample=False) -> dict:
total_iterations = 10**2
valid_iterations = total_iterations
metric = "relative"
sample_size = 10**3
delta = 0.05
size_array = [total_iterations, 2]
# [rows, columns]
if mc_dist.mc_enum == MCEnum.UNIFORM:
param_array = np.random.uniform(
low=0, high=mc_dist.param_list[0], size=size_array)
elif mc_dist.mc_enum == MCEnum.EXPONENTIAL:
param_array = np.random.exponential(
scale=mc_dist.param_list[0], size=size_array)
else:
raise NameError("Distribution parameter {0} is infeasible".format(
mc_dist.mc_enum))
res_array = np.empty([total_iterations, 3])
if sample:
res_array_sample = np.empty([total_iterations, 3])
for i in tqdm(range(total_iterations)):
setting = SingleServerPerform(
arr=DM1(lamb=param_array[i, 0]),
const_rate=ConstantRate(rate=param_array[i, 1]),
perform_param=perform_param)
theta_bounds = [(0.1, 4.0)]
bound_array = theta_bounds[:]
res_array[i, 0] = Optimize(setting=setting).grid_search(
bound_list=bound_array, delta=delta)
bound_array_power = theta_bounds[:]
bound_array_power.append((0.9, 4.0))
res_array[i, 1] = OptimizeNew(setting_new=setting).grid_search(
bound_list=bound_array_power, delta=delta)
if perform_param.perform_metric == PerformEnum.DELAY_PROB:
res_array[i, 2] = delay_prob_lower_exp_dm1_opt(
t=start_time,
delay=perform_param.value,
lamb=param_array[i, 0],
rate=param_array[i, 1])
if sample:
res_array_sample[i, 0] = res_array[i, 0]
res_array_sample[i, 1] = res_array[i, 1]
res_array_sample[i, 2] = delay_prob_sample_exp_dm1_opt(
t=start_time,
delay=perform_param.value,
lamb=param_array[i, 0],
rate=param_array[i, 1],
sample_size=sample_size)
if res_array[i, 0] > 1.0:
res_array[i, ] = nan
if sample:
res_array_sample[i, ] = nan
elif perform_param.perform_metric == PerformEnum.OUTPUT:
res_array[i, 2] = output_lower_exp_dm1_opt(
s=start_time,
delta_time=perform_param.value,
lamb=param_array[i, 0],
rate=param_array[i, 1])
else:
raise NameError("{0} is an infeasible performance metric".format(
perform_param.perform_metric))
if (res_array[i, 1] == inf or res_array[i, 2] == inf
or res_array[i, 0] == nan or res_array[i, 1] == nan
or res_array[i, 2] == nan):
res_array[i, ] = nan
res_array_sample[i, ] = nan
valid_iterations -= 1
# print("exponential results", res_array[:, 2])
res_dict = three_col_array_to_results(
arrival_enum=ArrivalEnum.DM1,
res_array=res_array,
valid_iterations=valid_iterations,
metric=metric)
res_dict.update({
"iterations": total_iterations,
"delta_time": perform_param.value,
"optimization": "grid_search",
"metric": "relative",
"MCDistribution": mc_dist.to_name(),
"MCParam": mc_dist.param_to_string()
})
res_dict_sample = three_col_array_to_results(
arrival_enum=ArrivalEnum.DM1,
res_array=res_array_sample,
valid_iterations=valid_iterations,
metric=metric)
res_dict_sample.update({
"iterations": total_iterations,
"delta_time": perform_param.value,
"optimization": "grid_search",
"metric": "relative",
"MCDistribution": mc_dist.to_name(),
"MCParam": mc_dist.param_to_string()
})
with open(
"lower_single_{0}_DM1_results_MC{1}_power_exp.csv".format(
perform_param.to_name(), mc_dist.to_name()), 'w') as csv_file:
writer = csv.writer(fileobj=csv_file)
for key, value in res_dict.items():
writer.writerow([key, value])
if sample:
with open(
"sample_single_{0}_DM1_results_MC{1}_power_exp.csv".format(
perform_param.to_name(), mc_dist.to_name()),
'w') as csv_file:
writer = csv.writer(fileobj=csv_file)
for key, value in res_dict_sample.items():
writer.writerow([key, value])
return res_dict
# for key, value in res_dict.items():
# writer.writerow([key, value])
#
# return res_dict
if __name__ == '__main__':
# TOTAL_ITERATIONS = 10**3
TOTAL_ITERATIONS = 200
METRIC = "relative"
DELTA = 0.05
START = 30
DELTA_TIME = 10
DELAY10 = PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=DELTA_TIME)
# DELTA_TIME = 4
# OUTPUT4 = PerformParameter(
# perform_metric=PerformEnum.OUTPUT, value=DELTA_TIME)
# LAMB = 1.0
# SERVICE_RATE = 1.2
#
# BOUND_LIST = [(0.05, 10.0)]
# BOUND_LIST_NEW = [(0.05, 10.0), (1.05, 20.0)]
# DELTA = 0.05
# PRINT_X = False
#
# CR_SERVER = ConstantRate(SERVICE_RATE)
#
def get_parameter_at_i(self, i: int) -> PerformParameter:
return PerformParameter(perform_metric=self.perform_metric,
value=self.values_list[i])
filename = name
filename += f"_time_{perform_param.to_name()}_{arrival_enum.name}" \
f"_MC{mc_dist.to_name()}_{opt_method.name}_util_{target_util}"
with open(filename + ".csv", 'w') as csv_file:
writer = csv.writer(csv_file)
for key, value in time_dict.items():
writer.writerow([key, value])
return time_dict
if __name__ == '__main__':
# DELAY_PROB10 = PerformParameter(perform_metric=PerformEnum.DELAY_PROB,
# value=10)
DELAY_6 = PerformParameter(perform_metric=PerformEnum.DELAY,
value=10**(-6))
COMMON_PERFORM_PARAM = DELAY_6
COMMON_OPTIMIZATION = OptMethod.GRID_SEARCH
COMMON_METRIC = ChangeEnum.RATIO_REF_NEW
TARGET_UTIL = 0.75
# MC_UNIF20 = MonteCarloDist(mc_enum=MCEnum.UNIFORM, param_list=[20.0])
MC_UNIF10 = MonteCarloDist(mc_enum=MCEnum.UNIFORM, param_list=[10.0])
PROCESS = ArrivalEnum.MD1
print(
csv_msob_fp_time(name="overlapping_tandem",
number_flows=3,
number_servers=3,
server_index=server_index)
results_df = pd.DataFrame(
{
"standard_bound": standard_bound,
"server_bound": server_bound,
"fp_bound": fp_bound
},
index=utilizations)
results_df = results_df[["standard_bound", "server_bound", "fp_bound"]]
return results_df
if __name__ == '__main__':
DELAY3 = PerformParameter(perform_metric=PerformEnum.DELAY, value=10**(-3))
# print(
# arrival_list_to_csv(
# prefix="overlapping_tandem_",
# data_frame_creator=overlapping_tandem_adjust_arr_df,
# list_arr_list=[[DM1(lamb=2.3),
# DM1(lamb=4.5),
# DM1(lamb=0.8)],
# [DM1(lamb=2.3),
# DM1(lamb=4.5),
# DM1(lamb=0.7)],
# [DM1(lamb=2.3),
# DM1(lamb=4.5),
# DM1(lamb=0.6)],
# [DM1(lamb=2.3),
def csv_tandem_compare_servers(
foi_arrival: ArrivalDistribution, cross_arrival: ArrivalDistribution,
foi_arrival2: ArrivalDistribution, cross_arrival2: ArrivalDistribution,
rate: float, max_servers: int, perform_param: PerformParameter,
opt_method: OptMethod, nc_analysis: NCAnalysis) -> pd.DataFrame:
"""Write dataframe results into a csv file.
Args:
foi_arrival: flow of interest's arrival distribution
foi_arrival2: competitor's flow of interest's arrival distribution
cross_arrival: distribution of cross arrivals
cross_arrival2: competitor's distribution of cross arrivals
rate: service rate of servers
max_servers: max number of servers in tandem
perform_param: performance parameter values
opt_method: optimization method
nc_analysis: Network Calculus analysis type
Returns:
csv file
"""
bounds = [0.0] * max_servers
bounds2 = [0.0] * max_servers
filename = "tandem_{0}".format(perform_param.to_name_value())
arr_list: List[ArrivalDistribution] = [foi_arrival]
arr_list2: List[ArrivalDistribution] = [foi_arrival2]
ser_list: List[ConstantRate] = []
for _i in range(max_servers):
print("current_number_servers {0}".format(_i + 1))
start = timer()
arr_list.append(cross_arrival)
arr_list2.append(cross_arrival2)
ser_list.append(ConstantRate(rate=rate))
bounds[_i], bounds2[_i] = tandem_compare(
arr_list=arr_list,
arr_list2=arr_list2,
ser_list=ser_list,
opt_method=opt_method,
perform_param=perform_param,
nc_analysis=nc_analysis)
end = timer()
print("duration: {0}".format(end - start))
filename += "_max" + str(max_servers) + "servers_" + foi_arrival.to_value(
) + "_rate=" + str(rate)
results_df = pd.DataFrame({
"bounds": bounds,
"bounds2": bounds2
},
index=range(1, max_servers + 1))
results_df = results_df[["bounds", "bounds2"]]
results_df.to_csv(
filename + '.csv', index=True, quoting=csv.QUOTE_NONNUMERIC)
return results_df
filename = name
filename += f"_results_{perform_param.to_name()}_{arrival_enum.name}_" \
f"MC{mc_dist.to_name()}_{opt_method.name}_" \
f"{compare_metric.name}_util_{target_util}"
with open(filename + ".csv", 'w') as csv_file:
writer = csv.writer(csv_file)
for key, value in res_dict.items():
writer.writerow([key, value])
return res_dict
if __name__ == '__main__':
COMMON_PERFORM_PARAM = PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=10)
# COMMON_PERFORM_PARAM = PerformParameter(perform_metric=PerformEnum.DELAY,
# value=1e-6)
COMMON_OPTIMIZATION = OptMethod.GRID_SEARCH
COMMON_METRIC = ChangeEnum.RATIO_REF_NEW
TARGET_UTIL = 0.7
# MC_UNIF20 = MonteCarloDist(mc_enum=MCEnum.UNIFORM, param_list=[20.0])
MC_UNIF10 = MonteCarloDist(mc_enum=MCEnum.UNIFORM, param_list=[10.0])
MC_EXP1 = MonteCarloDist(mc_enum=MCEnum.EXPONENTIAL, param_list=[1.0])
# ARRIVAL_PROCESSES = [
# ArrivalEnum.DM1, ArrivalEnum.MMOOFluid, ArrivalEnum.MD1
# ]
0], time_array[i,
1] = compute_overhead(setting=setting,
opt_method=opt_method,
number_l=1)
if i % floor(total_iterations / 10) == 0:
print(f"iteration {i} of {total_iterations}")
return time_array_to_results(arrival_enum=arrival_enum,
time_array=time_array,
number_servers=1,
time_ratio=time_ratio)
if __name__ == '__main__':
OUTPUT_TIME = PerformParameter(perform_metric=PerformEnum.OUTPUT, value=4)
COMMON_OPTIMIZATION = OptMethod.GRID_SEARCH
MC_UNIF20 = MonteCarloDist(mc_enum=MCEnum.UNIFORM, param_list=[20.0])
print(
mc_time_single(arrival_enum=ArrivalEnum.DM1,
perform_param=OUTPUT_TIME,
opt_method=COMMON_OPTIMIZATION,
mc_dist=MC_UNIF20))
print(
mc_time_single(arrival_enum=ArrivalEnum.MMOO,
perform_param=OUTPUT_TIME,
opt_method=COMMON_OPTIMIZATION,
grid_res = scipy.optimize.brute(func=helper_fun,
ranges=(slice(0.05, 20.0, 0.05), ),
full_output=True)
except (FloatingPointError, OverflowError):
return inf
if print_x:
print("grid search optimal x: ", grid_res[0].tolist())
return grid_res[1]
if __name__ == '__main__':
DELAY_VAL = 5
DELAY5 = PerformParameter(perform_metric=PerformEnum.DELAY_PROB,
value=DELAY_VAL)
DELAY_PROB_VAL = 10**(-5)
DELAY_PROB6 = PerformParameter(perform_metric=PerformEnum.DELAY,
value=DELAY_PROB_VAL)
NUMBER_AGGREGATIONS = 4
RHO_SINGLE = 1.0
SIGMA_SINGLE = 8.0
SERVICE_RATE = 6.5
BOUND_LIST = [(0.05, 20.0)]
DELTA = 0.05
PRINT_X = True
def csv_fat_cross_param_power(arrival_enum: ArrivalEnum, number_servers: int,
perform_param: PerformParameter,
opt_method: OptMethod,
mc_dist: MonteCarloDist) -> dict:
"""Chooses parameters by Monte Carlo type random choice."""
total_iterations = 10**4
valid_iterations = total_iterations
metric = "relative"
size_array = [
total_iterations,
(arrival_enum.number_parameters() + 1) * number_servers
# const_rate has 1 parameter
]
# [rows, columns]
param_array = mc_enum_to_dist(mc_dist=mc_dist, size=size_array)
res_array = np.empty([total_iterations, 2])
# print(res_array)
for i in tqdm(range(total_iterations), total=total_iterations):
if arrival_enum == ArrivalEnum.DM1:
arrive_list = [
DM1(lamb=param_array[i, j]) for j in range(number_servers)
]
elif arrival_enum == ArrivalEnum.MD1:
arrive_list = [
MD1(lamb=param_array[i, j],
mu=1 / (param_array[i,
arrival_enum.number_parameters() *
number_servers + j]))
for j in range(number_servers)
]
elif arrival_enum == ArrivalEnum.MMOO:
arrive_list = [
MMOOFluid(mu=param_array[i, j],
lamb=param_array[i, number_servers + j],
burst=param_array[i, 2 * number_servers + j])
for j in range(number_servers)
]
elif arrival_enum == ArrivalEnum.EBB:
arrive_list = [
EBB(factor_m=param_array[i, j],
decay=param_array[i, number_servers + j],
rho_single=param_array[i, 2 * number_servers + j])
for j in range(number_servers)
]
elif arrival_enum == ArrivalEnum.MassOne:
arrive_list = [
LeakyBucketMassOne(sigma_single=param_array[i, j],
rho_single=param_array[i,
number_servers + j],
n=20) for j in range(number_servers)
]
# TODO: note that n is fixed
elif arrival_enum == ArrivalEnum.TBConst:
arrive_list = [
TokenBucketConstant(sigma_single=param_array[i, j],
rho_single=param_array[i,
number_servers + j],
n=1) for j in range(number_servers)
]
else:
raise NameError("Arrival parameter {0} is infeasible".format(
arrival_enum.name))
if arrival_enum == ArrivalEnum.MD1 or arrival_enum == ArrivalEnum.MM1:
service_list = [
ConstantRate(rate=1.0) for j in range(number_servers)
]
else:
service_list = [
ConstantRate(
rate=param_array[i,
arrival_enum.number_parameters() *
number_servers + j])
for j in range(number_servers)
]
setting = FatCrossPerform(arr_list=arrive_list,
ser_list=service_list,
perform_param=perform_param)
# print(res_array[i, ])
# standard_bound, new_bound = compute_improvement()
res_array[i, 0], res_array[i, 1] = compute_improvement(
setting=setting,
opt_method=opt_method,
number_l=number_servers - 1)
if perform_param.perform_metric == PerformEnum.DELAY_PROB:
if res_array[i, 1] > 1.0:
res_array[i, ] = nan
if res_array[i, 0] == nan or res_array[i, 1] == nan:
res_array[i, ] = nan
valid_iterations -= 1
res_dict = two_col_array_to_results(arrival_enum=arrival_enum,
param_array=param_array,
res_array=res_array,
number_servers=number_servers,
valid_iterations=valid_iterations,
metric=metric)
res_dict.update({
"iterations": total_iterations,
"T": perform_param.value,
"optimization": opt_method.name,
"metric": metric,
"MCDistribution": mc_dist.to_name(),
"MCParam": mc_dist.param_to_string(),
"number_servers": number_servers
})
with open(
"sim_{0}_{1}_results_MC{2}_{3}_{4}.csv".format(
perform_param.to_name(), arrival_enum.name, mc_dist.to_name(),
opt_method.name, metric), 'w') as csv_file:
writer = csv.writer(csv_file)
for key, value in res_dict.items():
writer.writerow([key, value])
return res_dict
if __name__ == '__main__':
from nc_arrivals.iid import DM1
from nc_operations.perform_enum import PerformEnum
from nc_server.constant_rate_server import ConstantRateServer
from optimization.optimize import Optimize
from msob_and_fp.optimize_fp_bound import OptimizeFPBound
from msob_and_fp.optimize_server_bound import OptimizeServerBound
# from optimization.function_optimizer import optimizer_perform
# DELAY_PROB_TIME = PerformParameter(
# perform_metric=PerformEnum.DELAY_PROB, value=6)
# DELAY_PROB_TIME = PerformParameter(perform_metric=PerformEnum.DELAY_PROB,
# value=8)
DELAY_PROB_TIME = PerformParameter(perform_metric=PerformEnum.DELAY,
value=1E-3)
ARR_LIST = [DM1(lamb=2.3), DM1(lamb=4.5), DM1(lamb=1.7), DM1(lamb=4.5)]
SER_LIST = [
ConstantRateServer(rate=1.2),
ConstantRateServer(rate=6.2),
ConstantRateServer(rate=7.3),
ConstantRateServer(rate=6.2)
]
if ARR_LIST[0].is_discrete() is False:
raise ValueError("Distribution must be discrete")
RANGES = [slice(0.1, 10.0, 0.1)]
RANGES_2 = [slice(0.1, 10.0, 0.1), slice(1.1, 10.0, 0.1)]
number_servers=number_servers,
time_ratio=time_ratio))
filename = (f"time_{perform_param.to_name()}_{arrival_enum.name}"
f"_{opt_method.name}.csv")
with open(filename, 'w') as csv_file:
writer = csv.writer(csv_file)
for key, value in time_ratio.items():
writer.writerow([key, value])
return time_ratio
if __name__ == '__main__':
DELAY_PROB = PerformParameter(perform_metric=PerformEnum.DELAY_PROB,
value=4)
COMMON_OPTIMIZATION = OptMethod.PATTERN_SEARCH
MC_UNIF20 = MonteCarloDist(mc_enum=MCEnum.UNIFORM, param_list=[20.0])
list_number_servers1 = [2, 4, 6, 8, 10, 12]
print(
csv_fat_cross_time(arrival_enum=ArrivalEnum.DM1,
list_number_servers=list_number_servers1,
perform_param=DELAY_PROB,
opt_method=COMMON_OPTIMIZATION,
mc_dist=MC_UNIF20,
target_util=0.5))
def fat_cross_perform_df(arr_list: List[ArrivalDistribution],
ser_list: List[ConstantRateServer],
opt_method: OptMethod,
perform_param_list: PerformParamList) -> pd.DataFrame:
"""Compute delay standard_bound for T in T_list and write into dataframe.
Args:
arr_list: Arrival object list
ser_list: Service object list
opt_method: PS or GS
perform_param_list: list of performance parameter values
Returns:
dataframe
"""
delta_val = 0.05
one_param_bounds = [(delta_val, 10.0)]
standard_bound = [0.0] * len(perform_param_list)
h_mit_bound = [0.0] * len(perform_param_list)
print_x = False
fat_cross_setting = FatCrossPerform(
arr_list=arr_list,
ser_list=ser_list,
perform_param=PerformParameter(perform_metric=PerformEnum.DELAY_PROB,
value=0))
print(f"utilization: {fat_cross_setting.approximate_utilization()}")
print()
for i in range(len(perform_param_list)):
fat_cross_setting = FatCrossPerform(
arr_list=arr_list,
ser_list=ser_list,
perform_param=perform_param_list.get_parameter_at_i(i))
if opt_method == OptMethod.GRID_SEARCH:
standard_bound[i] = Optimize(setting=fat_cross_setting,
number_param=1,
print_x=print_x).grid_search(
grid_bounds=one_param_bounds,
delta=delta_val)
h_mit_bound[i] = OptimizeMitigator(
setting_h_mit=fat_cross_setting,
number_param=2,
print_x=print_x).grid_search(grid_bounds=[(delta_val, 10.0),
(1 + delta_val, 8.0)
],
delta=delta_val)
elif opt_method == OptMethod.PATTERN_SEARCH:
standard_bound[i] = Optimize(setting=fat_cross_setting,
number_param=1,
print_x=print_x).pattern_search(
start_list=[0.5],
delta=3.0,
delta_min=0.01)
h_mit_bound[i] = OptimizeMitigator(setting_h_mit=fat_cross_setting,
number_param=2,
print_x=print_x).pattern_search(
start_list=[0.5, 2.0],
delta=3.0,
delta_min=0.01)
else:
raise NotImplementedError(f"Optimization parameter {opt_method} "
f"is infeasible")
results_df = pd.DataFrame(
{
"standard_bound": standard_bound,
"h_mit_bound": h_mit_bound
},
index=perform_param_list.values_list)
return results_df
def csv_single_param_exp(start_time: int,
delay: int,
mc_dist: MonteCarloDist,
target_util: float,
total_iterations: int,
sample=False) -> dict:
valid_iterations = total_iterations
metric = ChangeEnum.RATIO_REF_NEW
sample_size = 10**2
delta = 0.05
size_array = [total_iterations, 2]
# [rows, columns]
if mc_dist.mc_enum == MCEnum.UNIFORM:
param_array = np.random.uniform(low=0,
high=mc_dist.param_list[0],
size=size_array)
elif mc_dist.mc_enum == MCEnum.EXPONENTIAL:
param_array = np.random.exponential(scale=mc_dist.param_list[0],
size=size_array)
else:
raise NameError(
f"Distribution parameter {mc_dist.mc_enum} is infeasible")
res_array = np.empty([total_iterations, 2])
res_array_sample = np.empty([total_iterations, 2])
for i in tqdm(range(total_iterations)):
single_setting = SingleServerMitPerform(
arr_list=[DM1(lamb=param_array[i, 0])],
server=ConstantRateServer(rate=param_array[i, 1]),
perform_param=PerformParameter(
perform_metric=PerformEnum.DELAY_PROB, value=delay))
computation_necessary = True
# print(res_array[i, ])
if target_util > 0.0:
util = single_setting.approximate_utilization()
if util < target_util or util > 1:
res_array[i, ] = np.nan
computation_necessary = False
if computation_necessary:
theta_bounds = [(0.1, 4.0)]
res_array[i, 0] = Optimize(setting=single_setting,
number_param=1).grid_search(
grid_bounds=theta_bounds,
delta=delta)
res_array[i,
1] = delay_prob_lower_exp_dm1_opt(t=start_time,
delay=delay,
lamb=param_array[i, 0],
rate=param_array[i, 1])
if sample:
res_array_sample[i, 1] = delay_prob_sample_exp_dm1_opt(
t=start_time,
delay=delay,
lamb=param_array[i, 0],
rate=param_array[i, 1],
sample_size=sample_size)
if res_array[i, 0] > 1.0:
res_array[i, ] = np.nan
if sample:
res_array_sample[i, ] = np.nan
if np.isnan(res_array[i, 0]) or np.isnan(res_array[i, 1]):
res_array[i, ] = np.nan
res_array_sample[i, ] = np.nan
valid_iterations -= 1
# print("exponential results", res_array[:, 2])
res_dict = two_col_array_to_results(arrival_enum=ArrivalEnum.DM1,
param_array=param_array,
res_array=res_array,
number_servers=1,
valid_iterations=valid_iterations,
compare_metric=metric)
res_dict.update({
"iterations": total_iterations,
"delta_time": delay,
"optimization": "grid_search",
"metric": metric.name,
"MCDistribution": mc_dist.to_name(),
"MCParam": mc_dist.param_to_string()
})
if sample:
res_array_sample[:, 0] = res_array[:, 0]
res_dict_sample = two_col_array_to_results(
arrival_enum=ArrivalEnum.DM1,
param_array=param_array,
res_array=res_array_sample,
number_servers=1,
valid_iterations=valid_iterations,
compare_metric=metric)
res_dict_sample.update({
"iterations": total_iterations,
"delta_time": delay,
"optimization": "grid_search",
"metric": metric.name,
"MCDistribution": mc_dist.to_name(),
"MCParam": mc_dist.param_to_string()
})
suffix = f"single_DELAY_PROB_DM1_results" \
f"_MC{mc_dist.to_name()}_power_exp.csv"
with open("lower_" + suffix, 'w') as csv_file:
writer = csv.writer(csv_file)
for key, value in res_dict.items():
writer.writerow([key, value])
if sample:
with open("sample_" + suffix, 'w') as csv_file:
writer = csv.writer(csv_file)
for key, value in res_dict_sample.items():
writer.writerow([key, value])
return res_dict
from fat_tree.fat_cross_perform import FatCrossPerform
from nc_arrivals.arrival_distribution import ArrivalDistribution
from nc_arrivals.markov_modulated import MMOOFluid
from nc_arrivals.qt import DM1
from nc_operations.perform_enum import PerformEnum
from nc_service.constant_rate_server import ConstantRate
from optimization.optimize import Optimize
from optimization.optimize_new import OptimizeNew
from single_server.single_server_perform import SingleServerPerform
from utils.perform_parameter import PerformParameter
if __name__ == '__main__':
# Single server output calculation
print("Single Server Performance Bounds:\n")
OUTPUT_TIME6 = PerformParameter(perform_metric=PerformEnum.OUTPUT, value=6)
SINGLE_SERVER = SingleServerPerform(arr=DM1(lamb=1.0),
const_rate=ConstantRate(rate=10.0),
perform_param=OUTPUT_TIME6)
print(SINGLE_SERVER.bound(param_list=[0.1]))
print(SINGLE_SERVER.new_bound(param_l_list=[0.1, 2.7]))
print(
Optimize(SINGLE_SERVER, print_x=True,
show_warn=True).grid_search(bound_list=[(0.1, 5.0)],
delta=0.1))
print(
OptimizeNew(SINGLE_SERVER, print_x=True,
theta=param_l_list[0],
delta_time=self.perform_param.value,
l_power=param_l_list[1])
else:
raise NameError(f"{self.perform_param.perform_metric} is an"
f"infeasible performance metric")
def to_string(self) -> str:
return self.to_name() + "_" + self.arr.to_value(
) + "_" + self.ser.to_value() + self.perform_param.to_name_value()
if __name__ == '__main__':
EXP_ARRIVAL1 = DM1(lamb=1.0)
CONST_RATE16 = ConstantRate(rate=1.6)
OUTPUT_4 = PerformParameter(perform_metric=PerformEnum.OUTPUT, value=4)
EX_OUTPUT = SingleServerPerform(arr=EXP_ARRIVAL1,
const_rate=CONST_RATE16,
perform_param=OUTPUT_4)
print(EX_OUTPUT.bound(param_list=[0.5]))
print(EX_OUTPUT.new_bound(param_l_list=[0.5, 1.2]))
DELAY_PROB_4 = PerformParameter(perform_metric=PerformEnum.DELAY_PROB,
value=4)
EX_DELAY_PROB = SingleServerPerform(arr=EXP_ARRIVAL1,
const_rate=CONST_RATE16,
perform_param=DELAY_PROB_4)
print(EX_DELAY_PROB.bound(param_list=[0.5]))
print(EX_DELAY_PROB.new_bound(param_l_list=[0.5, 1.2]))
from h_mitigator.fat_cross_perform import FatCrossPerform
from nc_arrivals.arrival_distribution import ArrivalDistribution
from nc_arrivals.markov_modulated import MMOOCont
from nc_arrivals.iid import DM1
from nc_operations.single_server_perform import SingleServerPerform
from nc_operations.perform_enum import PerformEnum
from nc_server.constant_rate_server import ConstantRateServer
from optimization.optimize import Optimize
from optimization.sim_anneal_param import SimAnnealParams
from utils.perform_parameter import PerformParameter
if __name__ == '__main__':
print("Single Server Performance Bounds:\n")
DELAY_PROB8 = PerformParameter(perform_metric=PerformEnum.DELAY_PROB,
value=8)
SINGLE_SERVER = SingleServerPerform(foi=DM1(lamb=1.0),
server=ConstantRateServer(rate=10.0),
perform_param=DELAY_PROB8)
print(
Optimize(SINGLE_SERVER, number_param=1,
print_x=True).grid_search(grid_bounds=[(0.1, 5.0)],
delta=0.1))
SINGLE_SERVER2 = SingleServerPerform(foi=MMOOCont(mu=0.7,
lamb=0.4,
peak_rate=1.2),
server=ConstantRateServer(rate=1.0),
perform_param=DELAY_PROB8)
