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 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.
"""
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