def get_invocations(benchmark: Benchmark):
"""
Returns a list of invocations that invoke the tool for the given benchmark.
It can be assumed that the current directory is the directory from which execute_invocations.py is executed.
For QCOMP 2020, this should return a list of invocations for all tracks in which the tool can take part. For each track an invocation with default settings has to be provided and in addition, an optimized setting (e.g., the fastest engine and/or solution technique for this benchmark) can be specified. Only information about the model type, the property type and the state space size are allowed to be used to tweak the parameters.
If this benchmark is not supported, an empty list has to be returned.
"""
if not is_benchmark_supported(benchmark):
return []
short = benchmark.get_model_short_name()
prop = benchmark.get_property_name()
prop_type = benchmark.get_short_property_type()
params = benchmark.get_parameter_values_string()
instance = short + "." + params
size = benchmark.get_num_states_tweak()
invocations = []
benchmark_settings = "--props " + benchmark.get_property_name()
if benchmark.get_open_parameter_def_string() != "":
benchmark_settings += " -E " + benchmark.get_open_parameter_def_string(
)
default_base = "modes/modest modes --unsafe --max-run-length 0 " + benchmark.get_janifilename(
) + " " + benchmark_settings + " -O out.txt Minimal"
#
# Track "probably-epsilon-correct"
#
precision = "5e-2"
default_cmd = default_base + " --width $PRECISION --relative-width"
if benchmark.is_dtmc() or benchmark.is_ctmc():
add_invocations(invocations, "probably-epsilon-correct",
default_cmd.replace("$PRECISION", precision))
#
# Track "often-epsilon-correct"
#
precision = "1e-3"
if benchmark.is_dtmc() or benchmark.is_ctmc():
add_invocations(invocations, "often-epsilon-correct",
default_cmd.replace("$PRECISION", precision))
#
# Track "often-epsilon-correct-10-min"
#
if benchmark.is_dtmc() or benchmark.is_ctmc():
default_cmd = default_base + " -N 2147483647"
else:
default_cmd = default_base + " --width 2e-2 --relative-width --lss Interruptible 1000000 -L 1000"
if benchmark.is_pta():
default_cmd += " --digital-clocks"
add_invocations(invocations, "often-epsilon-correct-10-min", default_cmd)
#
# Done
#
return invocations
def tweak_memory(benchmark : Benchmark): short = benchmark.get_model_short_name() params = benchmark.get_parameter_values_string() instance = short + "." + params prop = benchmark.get_property_name() full = short + "." + params + "." + prop tweak = "" if(instance == "beb.5-16-15" or instance == "coupon.15-4-5" or short == "egl" or short == "exploding-blocksworld" or short == "triangle-tireworld" or instance == "philosophers.20-1" or instance == "pnueli-zuck.10" or short == "rabin" or short == "tireworld" or short == "ftpp" or instance == "hecs.3-2" ): tweak += " -S Hybrid --store-compress None" else: tweak += " -S Memory" return tweak
def is_benchmark_supported(benchmark: Benchmark):
"""returns True if the provided benchmark is supported by the tool and if the given benchmark should appear on the generated benchmark list"""
short = benchmark.get_model_short_name()
prop = benchmark.get_property_name()
prop_type = benchmark.get_short_property_type()
if (short == "bluetooth" # multiple initial states
or short == "herman" # multiple initial states
or short == "oscillators" # model file too large, cannot be parsed
or short == "repudiation_malicious" # open clock constraints
):
return False
return True
def is_benchmark_supported(benchmark : Benchmark): """returns True if the provided benchmark is supported by the tool and if the given benchmark should appear on the generated benchmark list""" short = benchmark.get_model_short_name() prop = benchmark.get_property_name() if(short == "bluetooth" # multiple initial states or short == "cluster" and prop == "below_min" # bounded expected-reward property or short == "herman" # multiple initial states or short == "kanban" and prop == "throughput" # something's wrong with our algorithm here, didn't manage to investigate before deadline or short == "mapk_cascade" and prop == "reactions" # bounded expected-reward property or short == "oscillators" # model file too large, cannot be parsed or short == "repudiation_malicious" # open clock constraints or short == "resource-gathering" and prop == "expgold" # bounded expected-reward property or short == "resource-gathering" and prop == "prgoldgem" # unsupported property or short == "sms" and prop == "Unavailability" # Zeno MA model: not supported for long-run average rewards ): return False return True
def tweak(benchmark : Benchmark, specific_cmd): short = benchmark.get_model_short_name() params = benchmark.get_parameter_values_string() instance = short + "." + params prop = benchmark.get_property_name() full = short + "." + params + "." + prop tweak = "" if(full == "elevators.b-11-9.goal" or full == "rectangle-tireworld.11.goal" or full == "zenotravel.4-2-2.goal" or full == "firewire-pta.30-5000.eventually" ) and "--p1" not in specific_cmd: tweak += " --p1" elif(full == "firewire.false-36-800.deadline" ) and "--p0" not in specific_cmd: tweak += " --p0" return specific_cmd + tweak
def is_benchmark_supported(benchmark: Benchmark, trackId):
"""returns True if the provided benchmark is supported by the tool and if the given benchmark should appear on the generated benchmark list"""
if benchmark.is_pta() and benchmark.is_prism():
# Some PTAs from Prism are not supported because either
# modest can't apply digital clocks semantic due to open constraints, or
# modest puts the time as branch-rewards on the models, which are not supported for time-bounded properties
if benchmark.get_model_short_name() in [
"firewire-pta", "zeroconf-pta"
]:
return "time-bounded" not in benchmark.get_property_type()
else:
return False
if benchmark.is_prism_inf() and benchmark.is_ctmc():
# Storm does not support the CTMCs with infinite state-spaces
return False
# Time bounded queries on continuous time models can not be solved exactly
if trackId in ["correct", "floating-point-correct"]:
if "time-bounded" in benchmark.get_property_type() and (
benchmark.is_ma() or benchmark.is_ctmc()):
return False
return True
def is_benchmark_supported(benchmark: Benchmark):
"""
Returns True if the provided benchmark is supported by the tool and
if the given benchmark should appear on the generated benchmark list
"""
# DFTRES only supports Markovian models with purely spurious nondeterminism
if not (benchmark.is_dtmc() or benchmark.is_ctmc() or benchmark.is_ma()):
return False
# User-defined functions (the "call" JANI operator) are not supported
if "functions" in benchmark.get_jani_features():
return False
# Only time-accumulating or time-instant reward queries
supported_queries = [
"prob-reach", "prob-reach-time-bounded", "steady-state-prob"
]
if not benchmark.get_property_type() in supported_queries:
return False
# No support for real variables yet
real_vars = [ v for v in benchmark.load_jani_file()["variables"] \
if v["type"] == "real"]
if 0 < len(real_vars):
return False
# Some MAs have not-obviously-spurious nondeterminism and can't be simulated
unsupported_models = [
"bitcoin-attack",
]
# The arithmetic operations of some models aren't supported
unsupported_models += [
"majority", "philosophers", "speed-ind", "dpm", "readers-writers"
]
if benchmark.get_model_short_name() in unsupported_models:
return False
# All other models are supported
if ONLY_QCOMP_2020_BENCHMARKS:
return benchmark.get_identifier() in QComp2020_benchmarks
else:
return True
def get_invocations(benchmark: Benchmark):
"""
Returns a list of invocations that invoke the tool for the given benchmark.
It can be assumed that the current directory is the directory from which execute_invocations.py is executed.
For QCOMP 2020, this should return a list of invocations for all tracks in which the tool can take part. For each track an invocation with default settings has to be provided and in addition, an optimized setting (e.g., the fastest engine and/or solution technique for this benchmark) can be specified. Only information about the model type, the property type and the state space size are allowed to be used to tweak the parameters.
If this benchmark is not supported, an empty list has to be returned.
"""
# decide whether we want to use storm-dft (which supports galileo fault trees without repair)
use_storm_dft = False
if benchmark.is_galileo():
has_repair = False
for p in benchmark.get_file_parameters():
if p["name"] == "R" and p["value"] == True:
has_repair = True
if not has_repair:
use_storm_dft = True
# Gather the precision settings for the corresponding track
track_settings = dict()
track_comments = dict()
track_settings['correct'] = ' --exact '
track_comments['correct'] = 'Use exact arithmethic with rationals.'
track_settings[
'floating-point-correct'] = ' --exact floats --general:precision 1e-20 '
track_comments[
'floating-point-correct'] = 'Use exact arithmethic with floats. The precision needs to be set to increase precision when printing the result to stdout '
track_settings['epsilon-correct'] = ' --sound --precision 1e-6 '
track_comments['epsilon-correct'] = 'Use sound model checking methods.'
track_settings['probably-epsilon-correct'] = ' --sound --precision 5e-2 '
track_comments['probably-epsilon-correct'] = 'Use sound model checking.'
track_settings['often-epsilon-correct'] = ' --timebounded:precision 1e-3 '
track_comments[
'often-epsilon-correct'] = 'Use potentially unsound but fast solution methods. Use default precision (1e-6) everywhere except for timebounded queries, for which solution methods give epsilon guarantees.'
track_settings[
'often-epsilon-correct-10-min'] = ' --signal-timeout 60 --general:precision 1e-12 --gmm++:precision 1e-12 --native:precision 1e-12 --minmax:precision 1e-12 --timebounded:precision 1e-6 ' + (
"" if use_storm_dft else "--lra:precision 1e-12 ")
track_comments[
'often-epsilon-correct-10-min'] = 'Only force termination 60 seconds after receiving SIGTERM. Use potentially unsound but fast solution methods. Take a high precision to make sure that we make use of the 10 minutes. Time bounded queries can not be answered that precisely due to numerics.'
invocations = []
for trackId in track_settings:
if not is_benchmark_supported(benchmark, trackId):
continue
# Check whether this is a job for storm-dft
if use_storm_dft:
# We now have to obtain the correct property.
# Unfortunately, this is necessary because the gallileo files do not contain any information of the property
# The code below might easily break if we pick a different benchmark set
benchmark_settings = "--dftfile {} ".format(
benchmark.get_galileo_filename())
if benchmark.is_time_bounded_probabilistic_reachability():
time_bound = 1
for p in benchmark.get_parameters():
if p["name"] == "TIME_BOUND":
time_bound = p["value"]
benchmark_settings += "--timebound {} --max".format(time_bound)
elif benchmark.is_unbounded_expected_time():
benchmark_settings += "--expectedtime --min"
benchmark_settings += track_settings[trackId]
default_inv = Invocation()
default_inv.track_id = trackId
default_inv.identifier = "default"
default_inv.note = "Use Storm-dft with the requested property. " + track_comments[
trackId]
default_inv.add_command(
"~/storm/build/bin/storm-dft {}".format(benchmark_settings))
invocations.append(default_inv)
continue # with next trackId
# Gather options that are needed for this particular benchmark for any invocation of Storm
preprocessing_steps = []
benchmark_settings = ""
if (benchmark.is_prism()
or benchmark.is_prism_ma()) and not benchmark.is_pta():
benchmark_settings = "--prism {} --prop {} {}".format(
benchmark.get_prism_program_filename(),
benchmark.get_prism_property_filename(),
benchmark.get_property_name())
if benchmark.get_open_parameter_def_string() != "":
benchmark_settings += " --constants {}".format(
benchmark.get_open_parameter_def_string())
if benchmark.is_ctmc():
benchmark_settings += " --prismcompat"
else:
# For jani input, it might be the case that preprocessing is necessary using moconv
moconv_options = []
features = benchmark.get_jani_features()
for f in [
"arrays", "derived-operators", "functions",
"state-exit-rewards"
]:
if f in features: features.remove(f)
if "nondet-selection" in features:
moconv_options.append("--remove-disc-nondet")
features.remove("nondet-selection")
if len(features) != 0:
print("Unsupported jani feature(s): {}".format(features))
if benchmark.is_pta():
moconv_options.append("--digital-clocks")
if benchmark.get_model_short_name() == "wlan-large":
# This is actually a stochastic timed automaton. Distributions have to be unrolled first
moconv_options.append(" --unroll-distrs")
if len(moconv_options) != 0:
preprocessing_steps.append(
"~/modest/modest convert {} {} --output {} --overwrite".
format(benchmark.get_janifilename(),
" ".join(moconv_options),
"converted_" + benchmark.get_janifilename()))
if benchmark.get_open_parameter_def_string() != "":
preprocessing_steps[-1] += " --experiment {}".format(
benchmark.get_open_parameter_def_string())
benchmark_settings = "--jani {} --janiproperty {}".format(
"converted_" + benchmark.get_janifilename(),
benchmark.get_property_name())
else:
benchmark_settings = "--jani {} --janiproperty {}".format(
benchmark.get_janifilename(),
benchmark.get_property_name())
if benchmark.get_open_parameter_def_string() != "":
benchmark_settings += " --constants {}".format(
benchmark.get_open_parameter_def_string())
benchmark_settings += track_settings[trackId]
benchmark_settings += " --ddlib sylvan --sylvan:maxmem 6114 --sylvan:threads 4"
benchmark_comment = "Use sylvan as library for Dds, restricted to 6GB memory and 4 threads. " + track_comments[
trackId]
# default settings
default_inv = Invocation()
default_inv.identifier = "default"
# default_inv.note = benchmark_comment
default_inv.track_id = trackId
for prep in preprocessing_steps:
default_inv.add_command(prep)
default_inv.add_command(
"~/storm/build/bin/storm {}".format(benchmark_settings))
invocations.append(default_inv)
# specific settings
# Storm-static selects for each benchmark the best config among sparse, hybrid, ddbisim and exact (same for all tracks)
# We obtained this via previous experiments:
best_configs = dict()
best_configs["beb.4-8-7.LineSeized"] = "hybrid"
best_configs["beb.5-16-15.LineSeized"] = "N/A"
best_configs["bitcoin-attack.20-6.P_MWinMax"] = "ddbisim"
best_configs["bluetooth.1.time"] = "ddbisim"
best_configs["cabinets.3-2-true.Unavailability"] = "sparse"
best_configs["cabinets.3-2-true.Unreliability"] = "ddbisim"
best_configs["cluster.128-2000-20.premium_steady"] = "sparse"
best_configs["cluster.128-2000-20.qos1"] = "hybrid"
best_configs["cluster.64-2000-20.below_min"] = "hybrid"
best_configs["consensus.4-4.disagree"] = "sparse"
best_configs["consensus.4-4.steps_min"] = "sparse"
best_configs["consensus.6-2.disagree"] = "ddbisim"
best_configs["consensus.6-2.steps_min"] = "ddbisim"
best_configs["coupon.15-4-5.collect_all_bounded"] = "ddbisim"
best_configs["coupon.15-4-5.exp_draws"] = "ddbisim"
best_configs["coupon.9-4-5.collect_all_bounded"] = "ddbisim"
best_configs["coupon.9-4-5.exp_draws"] = "ddbisim"
best_configs["crowds.5-20.positive"] = "ddbisim"
best_configs["crowds.6-20.positive"] = "ddbisim"
best_configs["csma.3-4.all_before_max"] = "hybrid"
best_configs["csma.3-4.time_max"] = "hybrid"
best_configs["csma.4-2.all_before_max"] = "hybrid"
best_configs["csma.4-2.time_max"] = "hybrid"
best_configs["dpm.4-8-5.PmaxQueuesFullBound"] = "N/A"
best_configs["dpm.6-6-5.PminQueue1Full"] = "sparse"
best_configs["eajs.5-250-11.ExpUtil"] = "ddbisim"
best_configs["eajs.6-300-13.ExpUtil"] = "ddbisim"
best_configs["echoring.100.MaxOffline1"] = "sparse"
best_configs["egl.10-2.messagesB"] = "ddbisim"
best_configs["egl.10-2.unfairA"] = "hybrid"
best_configs["egl.10-8.messagesB"] = "ddbisim"
best_configs["egl.10-8.unfairA"] = "hybrid"
best_configs["elevators.b-11-9.goal"] = "sparse"
best_configs["embedded.8-12.actuators"] = "exact"
best_configs["embedded.8-12.up_time"] = "exact"
best_configs["exploding-blocksworld.10.goal"] = "N/A"
best_configs["firewire-pta.30-5000.eventually"] = "sparse"
best_configs["firewire.false-36-800.deadline"] = "ddbisim"
best_configs["fms.8.productivity"] = "N/A"
best_configs["ftpp.2-2-true.Unavailability"] = "N/A"
best_configs["ftwc.8-5.TimeMax"] = "sparse"
best_configs["ftwc.8-5.TimeMin"] = "sparse"
best_configs["haddad-monmege.100-0.7.exp_steps"] = "exact"
best_configs["haddad-monmege.100-0.7.target"] = "exact"
best_configs["hecs.false-1-1.Unreliability"] = "sparse"
best_configs["hecs.false-2-2.Unreliability"] = "sparse"
best_configs["hecs.false-3-2.Unreliability"] = "N/A"
best_configs["herman.15.steps"] = "ddbisim"
best_configs["kanban.5.throughput"] = "hybrid"
best_configs["majority.2100.change_state"] = "sparse"
best_configs["mapk_cascade.4-30.activated_time"] = "sparse"
best_configs["mapk_cascade.4-30.reactions"] = "hybrid"
best_configs["nand.40-4.reliable"] = "hybrid"
best_configs["nand.60-4.reliable"] = "hybrid"
best_configs[
"oscillators.8-10-0.1-1-0.1-1.0.power_consumption"] = "sparse"
best_configs["oscillators.8-10-0.1-1-0.1-1.0.time_to_synch"] = "sparse"
best_configs["pacman.100.crash"] = "hybrid"
best_configs["pacman.60.crash"] = "hybrid"
best_configs["philosophers.16-1.MaxPrReachDeadlock"] = "hybrid"
best_configs["philosophers.16-1.MaxPrReachDeadlockTB"] = "hybrid"
best_configs["philosophers.16-1.MinExpTimeDeadlock"] = "hybrid"
best_configs["philosophers.20-1.MaxPrReachDeadlock"] = "hybrid"
best_configs["philosophers.20-1.MaxPrReachDeadlockTB"] = "N/A"
best_configs["philosophers.20-1.MinExpTimeDeadlock"] = "N/A"
best_configs["pnueli-zuck.10.live"] = "hybrid"
best_configs["pnueli-zuck.5.live"] = "hybrid"
best_configs["polling.18-16.s1_before_s2"] = "hybrid"
best_configs["rabin.10.live"] = "hybrid"
best_configs["readers-writers.40.exp_time_many_requests"] = "sparse"
best_configs["readers-writers.40.prtb_many_requests"] = "hybrid"
best_configs["rectangle-tireworld.11.goal"] = "exact"
best_configs["resource-gathering.1300-100-100.expgold"] = "ddbisim"
best_configs["resource-gathering.1300-100-100.expsteps"] = "sparse"
best_configs["resource-gathering.1300-100-100.prgoldgem"] = "hybrid"
best_configs["sms.3-true.Unavailability"] = "hybrid"
best_configs["sms.3-true.Unreliability"] = "sparse"
best_configs["speed-ind.2100.change_state"] = "sparse"
best_configs["stream.1000.exp_buffertime"] = "sparse"
best_configs["stream.1000.pr_underrun"] = "sparse"
best_configs["stream.1000.pr_underrun_tb"] = "sparse"
best_configs["tireworld.45.goal"] = "N/A"
best_configs["triangle-tireworld.441.goal"] = "N/A"
best_configs["vgs.5-10000.MaxPrReachFailedTB"] = "N/A"
best_configs["vgs.5-10000.MinExpTimeFailed"] = "N/A"
best_configs["wlan-large.2.E_or"] = "sparse"
best_configs["wlan-large.2.P_max"] = "sparse"
best_configs["wlan.4-0.cost_min"] = "hybrid"
best_configs["wlan.4-0.sent"] = "hybrid"
best_configs["wlan.5-0.cost_min"] = "hybrid"
best_configs["wlan.5-0.sent"] = "hybrid"
best_configs["wlan.6-0.cost_min"] = "hybrid"
best_configs["wlan.6-0.sent"] = "hybrid"
best_configs["zenotravel.4-2-2.goal"] = "hybrid"
best_configs["zeroconf-pta.200.incorrect"] = "exact"
best_configs["zeroconf.1000-8-false.correct_max"] = "sparse"
best_configs["zeroconf.1000-8-false.correct_min"] = "sparse"
try:
config = best_configs[benchmark.get_identifier()]
except KeyError:
print(
"Unable to find best config for {}. Is this a new benchmark?".
format(benchmark.get_identifier()))
config = "N/A"
if config in ["N/A", "sparse"]:
# This is like the default config and thus does not need a rerun
continue
elif config == "hybrid":
benchmark_settings += " --engine hybrid"
elif config == "ddbisim":
benchmark_settings += " --engine dd-to-sparse --bisimulation"
elif config == "exact":
if trackId in ["correct", "floating-point-correct"]:
benchmark_settings += " --engine sparse"
else:
benchmark_settings += " --engine sparse --exact"
else:
assert False, "Unhandled config"
specific_inv = Invocation()
specific_inv.identifier = "specific"
specific_inv.track_id = trackId
for prep in preprocessing_steps:
specific_inv.add_command(prep)
specific_inv.add_command(
"~/storm/build/bin/storm {}".format(benchmark_settings))
invocations.append(specific_inv)
return invocations
def get_invocations(benchmark: Benchmark):
"""
Returns a list of invocations that invoke the tool for the given benchmark.
It can be assumed that the current directory is the directory from which execute_invocations.py is executed.
For QCOMP 2020, this should return a list of invocations for all tracks in which the tool can take part. For each track an invocation with default settings has to be provided and in addition, an optimized setting (e.g., the fastest engine and/or solution technique for this benchmark) can be specified. Only information about the model type, the property type and the state space size are allowed to be used to tweak the parameters.
If this benchmark is not supported, an empty list has to be returned.
"""
if not is_benchmark_supported(benchmark):
return []
# Gather options that are needed for this particular benchmark for any invocation of PRISM
benchmark_instance = get_prism_invocation_model_prop_instance(benchmark)
invocations = []
basic_args = "{}".format(prism_mem_args)
# epsilon-correct (all models but PTAs), default settings
if (benchmark.get_model_type() != "pta"):
# Use interval iteration generally (or uniformisation for time-bounded CTMCs)
default_args = "-ii"
# Choose engine heuristically
default_args += " -heuristic speed"
# Required precision (default anyway)
default_args += " -e 1e-6"
# Usual II settings when there is plenty of memory
default_args += " -ddextraactionvars 100"
# Increase maxiters (since QComp has a timeout anyway)
default_args += " -maxiters 1000000"
default_inv = Invocation()
default_inv.identifier = "default"
default_inv.track_id = "epsilon-correct"
default_inv.add_command(prism_bin + " " + basic_args + " " +
default_args + " " + benchmark_instance)
invocations.append(default_inv)
# epsilon-correct (all models but PTAs), specific settings
if (benchmark.get_model_type() != "pta"):
# Choose method/engine
# Use interval iteration generally (or uniformisation for time-bounded CTMCs)
if benchmark.get_model_short_name() == "haddad-monmege":
specific_args = "-exact"
elif (benchmark.get_num_states_tweak() == None
or benchmark.get_num_states_tweak() >= 20000000):
specific_args = "-ii -mtbdd"
else:
specific_args = "-ii -heuristic speed"
# Required precision (default anyway)
specific_args += " -e 1e-6"
# Usual II settings when there is plenty of memory
specific_args += " -ddextraactionvars 100"
# Increase maxiters (since QComp has a timeout anyway)
specific_args += " -maxiters 1000000"
specific_inv = Invocation()
specific_inv.identifier = "specific"
specific_inv.track_id = "epsilon-correct"
specific_inv.add_command(prism_bin + " " + basic_args + " " +
specific_args + " " + benchmark_instance)
invocations.append(specific_inv)
# often-epsilon-correct (all models), default settings
if (True):
# Choose engine heuristically
default_args = "-heuristic speed"
# Required precision (just use default 1e-6, as agreed for QComp'19)
default_args += " -e 1e-6"
# Increase maxiters (since QComp has a timeout anyway)
default_args += " -maxiters 1000000"
default_inv = Invocation()
default_inv.identifier = "default"
default_inv.track_id = "often-epsilon-correct"
default_inv.add_command(prism_bin + " " + basic_args + " " +
default_args + " " + benchmark_instance)
invocations.append(default_inv)
# often-epsilon-correct (all models), specific settings
if (True):
# Choose method/engine
if benchmark.get_model_short_name() == "haddad-monmege":
specific_args = "-exact"
elif (benchmark.get_num_states_tweak() == None
or benchmark.get_num_states_tweak() >= 20000000):
specific_args = "-mtbdd"
else:
specific_args = "-heuristic speed"
# Required precision (just use default 1e-6, as agreed for QComp'19)
specific_args += " -e 1e-6"
# Increase maxiters (since QComp has a timeout anyway)
specific_args += " -maxiters 1000000"
specific_inv = Invocation()
specific_inv.identifier = "specific"
specific_inv.track_id = "often-epsilon-correct"
specific_inv.add_command(prism_bin + " " + basic_args + " " +
specific_args + " " + benchmark_instance)
invocations.append(specific_inv)
# probably-epsilon-correct (all models but PTAs), default settings
if (benchmark.get_model_type() != "pta"):
# Use interval iteration generally (or uniformisation for time-bounded CTMCs)
if (benchmark.get_model_type() == "ctmc"
and benchmark.is_time_bounded_probabilistic_reachability()):
default_args = ""
else:
default_args = "-ii -e 5e-2"
# Choose engine heuristically
default_args += " -heuristic speed"
# Usual II settings when there is plenty of memory
default_args += " -ddextraactionvars 100"
# Increase maxiters (since QComp has a timeout anyway)
default_args += " -maxiters 1000000"
default_inv = Invocation()
default_inv.identifier = "default"
default_inv.track_id = "probably-epsilon-correct"
default_inv.add_command(prism_bin + " " + basic_args + " " +
default_args + " " + benchmark_instance)
invocations.append(default_inv)
# probably-epsilon-correct (all models but PTAs), specific settings
if (benchmark.get_model_type() != "pta"):
# Choose method/engine
# Use interval iteration generally (or uniformisation for time-bounded CTMCs)
if benchmark.get_model_short_name() == "haddad-monmege":
specific_args = "-exact"
elif (benchmark.get_model_type() == "ctmc"
and benchmark.is_time_bounded_probabilistic_reachability()):
specific_args = ""
elif (benchmark.get_num_states_tweak() == None
or benchmark.get_num_states_tweak() >= 20000000):
specific_args = "-ii -e 5e-2 -mtbdd"
else:
specific_args = "-ii -e 5e-2 -heuristic speed"
# Usual II settings when there is plenty of memory
specific_args += " -ddextraactionvars 100"
# Increase maxiters (since QComp has a timeout anyway)
specific_args += " -maxiters 1000000"
specific_inv = Invocation()
specific_inv.identifier = "specific"
specific_inv.track_id = "probably-epsilon-correct"
specific_inv.add_command(prism_bin + " " + basic_args + " " +
specific_args + " " + benchmark_instance)
invocations.append(specific_inv)
return invocations
def get_invocations(benchmark : Benchmark):
"""
Returns a list of invocations that invoke the tool for the given benchmark.
It can be assumed that the current directory is the directory from which execute_invocations.py is executed.
For QCOMP 2020, this should return a list of invocations for all tracks in which the tool can take part. For each track an invocation with default settings has to be provided and in addition, an optimized setting (e.g., the fastest engine and/or solution technique for this benchmark) can be specified. Only information about the model type, the property type and the state space size are allowed to be used to tweak the parameters.
If this benchmark is not supported, an empty list has to be returned.
"""
if not is_benchmark_supported(benchmark):
return []
short = benchmark.get_model_short_name()
prop = benchmark.get_property_name()
prop_type = benchmark.get_short_property_type()
params = benchmark.get_parameter_values_string()
instance = short + "." + params
size = benchmark.get_num_states_tweak()
invocations = []
benchmark_settings = ""
if benchmark.get_open_parameter_def_string() != "":
benchmark_settings += "-E " + benchmark.get_open_parameter_def_string() + " "
benchmark_settings += "--props " + benchmark.get_property_name()
default_base = "mcsta/modest mcsta " + benchmark.get_janifilename() + " " + benchmark_settings + " -O out.txt Minimal --unsafe --es"
specific_base = default_base + tweak_memory(benchmark) # specific settings
default_base += " -S Memory"
#
# Track "floating-point-correct"
#
skip = False
precision = "0"
default_cmd = default_base + " --no-partial-results"
specific_cmd = specific_base + " --no-partial-results"
if prop_type == "S" or (benchmark.is_ma() or benchmark.is_ctmc()) and benchmark.is_time_bounded_probabilistic_reachability():
# long-run average or time-bounded on MA/CTMC: no fp-exact algorithm available
skip = True
elif prop_type == "P":
# probabilistic reachability: try value iteration until fp-fixpoint
default_cmd += " --p0 --p1 --epsilon 0 --absolute-epsilon"
specific_cmd += " --p0 --p1 --epsilon 0 --absolute-epsilon"
elif prop_type == "E":
# expected reward: try value iteration until fp-fixpoint
default_cmd += " --epsilon 0 --absolute-epsilon"
specific_cmd += " --epsilon 0 --absolute-epsilon"
elif prop_type == "Pb":
# state elimination is fp-exact
default_cmd += " --reward-bounded-alg StateElimination"
if "-S Memory" in specific_cmd:
specific_cmd += " --reward-bounded-alg StateElimination"
else:
specific_cmd += " --epsilon 0 --absolute-epsilon"
if not skip:
add_invocations(invocations, "floating-point-correct", default_cmd, tweak(benchmark, specific_cmd))
#
# Track "epsilon-correct"
#
skip = False
precision = "1e-6"
default_cmd = default_base + " --no-partial-results"
specific_cmd = specific_base + " --no-partial-results"
if prop_type == "S":
# long-run average: default is the sound algorithm based on value iteration
default_cmd += " --width $PRECISION --relative-width"
specific_cmd += " --width $PRECISION --relative-width"
elif (benchmark.is_ma() or benchmark.is_ctmc()) and benchmark.is_time_bounded_probabilistic_reachability():
# time-bounded probability for CTMC and MA: default is sound Unif+
default_cmd += " --width $PRECISION --relative-width"
specific_cmd += " --width $PRECISION --relative-width"
elif benchmark.is_pta() and prop_type == "Pb":
# time-bounded reachability for PTA: state elimination recommended
default_cmd += " --reward-bounded-alg StateElimination"
if "-S Memory" in specific_cmd:
specific_cmd += " --reward-bounded-alg StateElimination"
else:
specific_cmd += " --alg IntervalIteration --width $PRECISION --relative-width"
elif prop_type == "Pb":
# reward-bounded probability: default is unsound VI, so need to change to II (SVI and OVI not yet implemented for this case)
default_cmd += " --alg IntervalIteration --width $PRECISION --relative-width"
specific_cmd += " --alg IntervalIteration --width $PRECISION --relative-width"
else:
# unbounded probability or expected reward: default is unsound VI, so need to change to OVI
default_cmd += " --alg OptimisticValueIteration --epsilon $PRECISION --width $PRECISION --relative-width"
specific_cmd += " --alg OptimisticValueIteration --epsilon $PRECISION --width $PRECISION --relative-width"
if prop_type == "P" and benchmark.is_dtmc() or benchmark.is_ctmc():
# for unbounded probabilities on DTMC and CTMC: use 0/1 preprocessing
default_cmd += " --p0 --p1"
specific_cmd += " --p0 --p1"
if not skip:
add_invocations(invocations, "epsilon-correct", default_cmd.replace("$PRECISION", precision), tweak(benchmark, specific_cmd).replace("$PRECISION", precision))
#
# Track "probably-epsilon-correct"
#
skip = False
precision = "5e-2"
if not skip:
add_invocations(invocations, "probably-epsilon-correct", default_cmd.replace("$PRECISION", precision), tweak(benchmark, specific_cmd).replace("$PRECISION", precision))
#
# Track "often-epsilon-correct"
#
skip = False
precision = "1e-3"
default_cmd = default_base + " --no-partial-results"
specific_cmd = specific_base + " --no-partial-results"
if prop_type == "S":
# long-run average: default is the sound algorithm based on value iteration
default_cmd += " --width $PRECISION --relative-width"
specific_cmd += " --width $PRECISION --relative-width"
elif (benchmark.is_ma() or benchmark.is_ctmc()) and benchmark.is_time_bounded_probabilistic_reachability():
# time-bounded probability for CTMC and MA: default is sound Unif+
default_cmd += " --width $PRECISION --relative-width"
specific_cmd += " --width $PRECISION --relative-width"
elif benchmark.is_pta() and prop_type == "Pb":
# time-bounded reachability for PTA: state elimination recommended
default_cmd += " --reward-bounded-alg StateElimination"
if "-S Memory" in specific_cmd:
specific_cmd += " --reward-bounded-alg StateElimination"
else:
specific_cmd += " --alg IntervalIteration --width $PRECISION --relative-width"
elif prop_type == "Pb":
# reward-bounded probability: default is unsound VI, which is okay here
pass
else:
# unbounded probability or expected reward: default is unsound VI, which is okay here
if prop_type == "P" and benchmark.is_dtmc() or benchmark.is_ctmc():
# for unbounded probabilities on DTMC and CTMC: use 0/1 preprocessing
default_cmd += " --p0 --p1"
specific_cmd += " --p0 --p1"
if not skip:
add_invocations(invocations, "often-epsilon-correct", default_cmd.replace("$PRECISION", precision), tweak(benchmark, specific_cmd).replace("$PRECISION", precision))
#
# Track "often-epsilon-correct-10-min"
#
skip = False
precision = "0" # so we just run for the full 10 minutes (or until we get an exact result)
default_cmd = default_base
specific_cmd = specific_base
if prop_type == "S":
skip = True
elif (benchmark.is_ma() or benchmark.is_ctmc()) and benchmark.is_time_bounded_probabilistic_reachability():
default_cmd += " --width 0"
specific_cmd += " --width 0"
elif prop_type == "Pb":
# reward-bounded reachability for DTMC, MDP, and PTA
default_cmd += " --reward-bounded-alg StateElimination"
if "-S Memory" in specific_cmd:
specific_cmd += " --reward-bounded-alg StateElimination"
else:
specific_cmd += " --epsilon 0"
else:
default_cmd += " --epsilon 0"
specific_cmd += " --epsilon 0"
if prop_type == "P" and benchmark.is_dtmc() or benchmark.is_ctmc():
# for unbounded probabilities on DTMC and CTMC: use 0/1 preprocessing
default_cmd += " --p0 --p1"
specific_cmd += " --p0 --p1"
if not skip:
add_invocations(invocations, "often-epsilon-correct-10-min", default_cmd, tweak(benchmark, specific_cmd))
#
# Done
#
return invocations
def get_invocations(benchmark: Benchmark):
"""
Returns a list of invocations that invoke the tool for the given benchmark.
It can be assumed that the current directory is the directory from which execute_invocations.py is executed.
For QCOMP 2020, this should return a list of invocations for all tracks in which the tool can take part. For each track an invocation with default settings has to be provided and in addition, an optimized setting (e.g., the fastest engine and/or solution technique for this benchmark) can be specified. Only information about the model type, the property type and the state space size are allowed to be used to tweak the parameters.
If this benchmark is not supported, an empty list has to be returned.
"""
# decide whether we want to use storm-dft (which supports galileo fault trees without repair)
use_storm_dft = False
if benchmark.is_galileo():
has_repair = False
for p in benchmark.get_file_parameters():
if p["name"] == "R" and p["value"] == True:
has_repair = True
if not has_repair:
use_storm_dft = True
# Gather the precision settings for the corresponding track
track_settings = dict()
track_comments = dict()
track_settings['correct'] = ' --exact '
track_comments['correct'] = 'Use exact arithmethic with rationals.'
track_settings[
'floating-point-correct'] = ' --exact floats --general:precision 1e-20 '
track_comments[
'floating-point-correct'] = 'Use exact arithmethic with floats. The precision needs to be set to increase precision when printing the result to stdout '
track_settings['epsilon-correct'] = ' --sound --precision 1e-6 '
track_comments['epsilon-correct'] = 'Use sound model checking methods.'
track_settings['probably-epsilon-correct'] = ' --sound --precision 5e-2 '
track_comments['probably-epsilon-correct'] = 'Use sound model checking.'
track_settings['often-epsilon-correct'] = ' --timebounded:precision 1e-3 '
track_comments[
'often-epsilon-correct'] = 'Use potentially unsound but fast solution methods. Use default precision (1e-6) everywhere except for timebounded queries, for which solution methods give epsilon guarantees.'
track_settings[
'often-epsilon-correct-10-min'] = ' --signal-timeout 60 --general:precision 1e-12 --gmm++:precision 1e-12 --native:precision 1e-12 --minmax:precision 1e-12 --timebounded:precision 1e-6 ' + (
"" if use_storm_dft else "--lra:precision 1e-12 ")
track_comments[
'often-epsilon-correct-10-min'] = 'Only force termination 60 seconds after receiving SIGTERM. Use potentially unsound but fast solution methods. Take a high precision to make sure that we make use of the 10 minutes. Time bounded queries can not be answered that precisely due to numerics.'
invocations = []
for trackId in track_settings:
if not is_benchmark_supported(benchmark, trackId):
continue
# Check whether this is a job for storm-dft
if use_storm_dft:
# We now have to obtain the correct property.
# Unfortunately, this is necessary because the gallileo files do not contain any information of the property
# The code below might easily break if we pick a different benchmark set
benchmark_settings = "--dftfile {} ".format(
benchmark.get_galileo_filename())
if benchmark.is_time_bounded_probabilistic_reachability():
time_bound = 1
for p in benchmark.get_parameters():
if p["name"] == "TIME_BOUND":
time_bound = p["value"]
benchmark_settings += "--timebound {} --max".format(time_bound)
elif benchmark.is_unbounded_expected_time():
benchmark_settings += "--expectedtime --min"
benchmark_settings += track_settings[trackId]
default_inv = Invocation()
default_inv.track_id = trackId
default_inv.identifier = "default"
default_inv.note = "Use Storm-dft with the requested property. " + track_comments[
trackId]
default_inv.add_command(
"~/storm/build/bin/storm-dft {}".format(benchmark_settings))
invocations.append(default_inv)
continue # with next trackId
# Gather options that are needed for this particular benchmark for any invocation of Storm
preprocessing_steps = []
benchmark_settings = ""
if (benchmark.is_prism()
or benchmark.is_prism_ma()) and not benchmark.is_pta():
benchmark_settings = "--prism {} --prop {} {}".format(
benchmark.get_prism_program_filename(),
benchmark.get_prism_property_filename(),
benchmark.get_property_name())
if benchmark.get_open_parameter_def_string() != "":
benchmark_settings += " --constants {}".format(
benchmark.get_open_parameter_def_string())
if benchmark.is_ctmc():
benchmark_settings += " --prismcompat"
else:
# For jani input, it might be the case that preprocessing is necessary using moconv
moconv_options = []
features = benchmark.get_jani_features()
for f in [
"arrays", "derived-operators", "functions",
"state-exit-rewards"
]:
if f in features: features.remove(f)
if "nondet-selection" in features:
moconv_options.append("--remove-disc-nondet")
features.remove("nondet-selection")
if len(features) != 0:
print("Unsupported jani feature(s): {}".format(features))
if benchmark.is_pta():
moconv_options.append("--digital-clocks")
if benchmark.get_model_short_name() == "wlan-large":
# This is actually a stochastic timed automaton. Distributions have to be unrolled first
moconv_options.append(" --unroll-distrs")
if len(moconv_options) != 0:
preprocessing_steps.append(
"~/modest/modest convert {} {} --output {} --overwrite".
format(benchmark.get_janifilename(),
" ".join(moconv_options),
"converted_" + benchmark.get_janifilename()))
if benchmark.get_open_parameter_def_string() != "":
preprocessing_steps[-1] += " --experiment {}".format(
benchmark.get_open_parameter_def_string())
benchmark_settings = "--jani {} --janiproperty {}".format(
"converted_" + benchmark.get_janifilename(),
benchmark.get_property_name())
else:
benchmark_settings = "--jani {} --janiproperty {}".format(
benchmark.get_janifilename(),
benchmark.get_property_name())
if benchmark.get_open_parameter_def_string() != "":
benchmark_settings += " --constants {}".format(
benchmark.get_open_parameter_def_string())
benchmark_settings += track_settings[trackId]
benchmark_settings += " --engine portfolio --ddlib sylvan --sylvan:maxmem 6114 --sylvan:threads 4"
benchmark_comment = "Use Storm with protfolio engine. Use sylvan as library for Dds, restricted to 6GB memory and 4 threads. " + track_comments[
trackId]
# default settings
default_inv = Invocation()
default_inv.identifier = "default"
# Apparently, a note is not needed
# default_inv.note = benchmark_comment
default_inv.track_id = trackId
for prep in preprocessing_steps:
default_inv.add_command(prep)
default_inv.add_command(
"~/storm/build/bin/storm {}".format(benchmark_settings))
invocations.append(default_inv)
# specific settings !!!!only information about model type, property type and state space size via benchmark.get_num_states_tweak() may be used for tweaking
# Omitted because there is no significant benefit
# if benchmark.get_num_states_tweak() is not None:
# specific_inv = Invocation()
# specific_inv.identifier = "specific"
# specific_inv.track_id = trackId
# for prep in preprocessing_steps:
# specific_inv.add_command(prep)
# specific_inv.add_command("~/storm/build/bin/storm {} --hints:states {}".format(benchmark_settings, benchmark.get_num_states_tweak()))
# invocations.append(specific_inv)
return invocations
