def invoke(self, env: Environment, replica: FunctionReplica, request: FunctionRequest):
# you would probably either create one simulator per function, or use a generalized simulator, this is just
# to demonstrate how the simulators are used to encapsulate simulator behavior.
logger.info('[simtime=%.2f] invoking function %s on node %s', env.now, request, replica.node.name)
# for full flexibility you decide the resources used
cpu_millis = replica.node.capacity.cpu_millis * 0.1
env.resource_state.put_resource(replica, 'cpu', cpu_millis)
node = replica.node
node.current_requests.add(request)
if replica.function.name == 'python-pi':
if replica.node.name.startswith('rpi3'): # those are nodes we created in basic.example_topology()
yield env.timeout(20) # invoking this function takes 20 seconds on a raspberry pi
else:
yield env.timeout(2) # invoking this function takes 2 seconds on all other nodes in the cluster
elif replica.function.name == 'resnet50-inference':
yield env.timeout(0.5) # invoking this function takes 500 ms
else:
yield env.timeout(0)
# also, you have to release them at the end
env.resource_state.remove_resource(replica, 'cpu', cpu_millis)
node.current_requests.remove(request)
def faas_idler(env: Environment,
inactivity_duration=300,
reconcile_interval=30):
"""
https://github.com/openfaas-incubator/faas-idler
https://github.com/openfaas-incubator/faas-idler/blob/master/main.go
default values:
https://github.com/openfaas-incubator/faas-idler/blob/668991c532156275993399ee79a297a4c2d651ec/docker-compose.yml
:param env: the faas environment
:param inactivity_duration: i.e. 15m (Golang duration)
:param reconcile_interval: i.e. 1m (default value)
:return: an event generator
"""
faas: FaasSystem = env.faas
while True:
yield env.timeout(reconcile_interval)
for deployment in faas.get_deployments():
if not deployment.scale_zero:
continue
for function in deployment.function_definitions.values():
if len(faas.get_replicas(function.name,
FunctionState.RUNNING)) == 0:
continue
idle_time = env.now - env.metrics.last_invocation[
function.name]
if idle_time >= inactivity_duration:
env.process(faas.suspend(function.name))
logger.debug('%.2f function %s has been idle for %.2fs',
env.now, function.name, idle_time)
def run(self, env: Environment):
# deploy functions
deployments = self.prepare_deployments()
for deployment in deployments:
yield from env.faas.deploy(deployment)
# block until replicas become available (scheduling has finished and replicas have been deployed on the node)
logger.info('waiting for replica')
yield env.process(env.faas.poll_available_replica('python-pi'))
yield env.process(env.faas.poll_available_replica('resnet50-inference'))
# run workload
ps = []
# execute 10 requests in parallel
logger.info('executing 10 python-pi requests')
for i in range(10):
ps.append(env.process(env.faas.invoke(FunctionRequest('python-pi'))))
logger.info('executing 10 resnet50-inference requests')
for i in range(10):
ps.append(env.process(env.faas.invoke(FunctionRequest('resnet50-inference'))))
# wait for invocation processes to finish
for p in ps:
yield p
def execute(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
# mock download, for actual network download simulation look at simulate_data_download
yield env.timeout(1)
# training
yield env.timeout(5)
# mock upload
yield env.timeout(1)
def solve(self,
env: Environment) -> Generator[simpy.events.Event, Any, Any]:
while self.running:
yield env.timeout(self.reconcile_interval)
yield from self.solver.solve(env)
# TODO remove when contention is implemented
self.stop()
def simulate_data_upload(env: Environment, replica: FunctionReplica):
node = replica.node.ether_node
func = replica
started = env.now
if 'data.skippy.io/sends-to-storage' not in func.pod.spec.labels:
return
# FIXME: storage
size = parse_size_string(
func.pod.spec.labels['data.skippy.io/sends-to-storage'])
path = func.pod.spec.labels['data.skippy.io/sends-to-storage/path']
storage_node_name = env.cluster.get_storage_nodes(path)[0]
logger.debug('%.2f replica %s uploading data %s to %s', env.now, node,
path, storage_node_name)
if storage_node_name == node.name:
# FIXME this is essentially a disk read and not a network connection
yield env.timeout(size / 1.25e+8) # 1.25e+8 = 1 GBit/s
return
storage_node = env.cluster.get_node(storage_node_name)
route = env.topology.route_by_node_name(node.name, storage_node.name)
flow = SafeFlow(env, size, route)
yield flow.start()
for hop in route.hops:
env.metrics.log_network(size, 'data_upload', hop)
env.metrics.log_flow(size, env.now - started, route.source,
route.destination, 'data_upload')
def teardown(self, env: Environment, replica: FunctionReplica):
# basic cpu usage, in %
env.resource_state.remove_resource(replica, 'cpu', 0.08)
# basic memory consumption, in MB
env.resource_state.remove_resource(replica, 'memory', 200)
yield env.timeout(0)
def startup(self, env: Environment, replica: FunctionReplica):
logger.info(
'[simtime=%.2f] starting up function replica for function %s',
env.now, replica.function.name)
# you could create a very fine-grained setup routines here
yield env.timeout(10) # simulate docker startup
def invoke(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
token = self.queue.request()
t_wait_start = env.now
yield token # wait for access
t_wait_end = env.now
t_fet_start = env.now
# because of GIL and Threads, we can easily estimate the additional time caused by concurrent requests to the
# same Function
factor = max(1, self.scale(self.queue.count, self.queue.capacity))
try:
fet = self.deployment.sample_fet(replica.node.name)
if fet is None:
logging.error(
f"FET for node {replica.node.name} for function {self.deployment.image} was not found"
)
raise ValueError(f'{replica.node.name}')
fet = float(fet) * factor
image = replica.pod.spec.containers[0].image
if 'preprocessing' in image or 'training' in image:
yield from simulate_data_download(env, replica)
start = env.now
# replica.node.current_requests.add(request)
call = FunctionCall(request, replica, start)
replica.node.all_requests.append(call)
yield env.timeout(fet)
# add degradation
end = env.now
degradation = replica.node.estimate_degradation(start, end)
delay = max(0, (fet * degradation) - fet)
yield env.timeout(delay)
if 'preprocessing' in image or 'training' in image:
yield from simulate_data_upload(env, replica)
t_fet_end = env.now
env.metrics.log_fet(request.name, replica.function.image,
replica.node.name, t_fet_start, t_fet_end,
t_wait_start, t_wait_end, degradation,
id(replica))
replica.node.set_end(request.request_id, end + delay)
# replica.node.current_requests.remove(request)
except KeyError:
pass
self.queue.release(token)
def invoke(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
# 1) get parameters of base distribution (ideal case)
# 2) check the utilization of the node the replica is running on
# 3) transform distribution parameters with degradation function depending on utilization
# 4) sample from that distribution
logger.info('invoking %s on %s (%d in parallel)', request.name,
replica.node.name, len(replica.node.current_requests))
yield env.timeout(1)
def __init__(self,
topology: Topology,
benchmark: Benchmark,
env: Environment = None,
timeout=None,
name=None):
self.env = env or Environment()
self.topology = topology
self.benchmark = benchmark
self.timeout = timeout
self.name = name
def run(self, env: Environment):
for deployment in self.deployments:
yield from env.faas.deploy(deployment)
for deployment in self.deployments:
yield env.process(env.faas.poll_available_replica(deployment.name))
ps = []
logging.info('executing requests')
for deployment in self.deployments:
try:
ia_generator = self.arrival_profiles[deployment.name]
if self.duration is None:
p = env.process(function_trigger(env, deployment, ia_generator, max_requests=1000))
else:
p = env.process(function_trigger(env, deployment, ia_generator))
ps.append(p)
except KeyError:
logging.warning('no arrival profile for deployment %s', deployment.name)
if self.duration is not None:
env.process(self.wait(env, ps))
yield from ps
def invoke(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
# you would probably either create one simulator per function, or use a generalized simulator, this is just
# to demonstrate how the simulators are used to encapsulate simulator behavior.
logger.info('[simtime=%.2f] invoking function %s on node %s', env.now,
request, replica.node.name)
if replica.function.name == 'python-pi':
if replica.node.name.startswith(
'rpi3'
): # those are nodes we created in basic.example_topology()
yield env.timeout(
20
) # invoking this function takes 20 seconds on a raspberry pi
else:
yield env.timeout(
2
) # invoking this function takes 2 seconds on all other nodes in the cluster
elif replica.function.name == 'resnet50-inference':
yield env.timeout(0.5) # invoking this function takes 500 ms
else:
yield env.timeout(0)
def pull(env: Environment, image_str: str, node: Node):
"""
Simulate a docker pull command of the given image on the given node.
:param env: the simulation environment
:param image_str: the name of the image (<repository[:tag]>)
:param node: the node on which to run the pull command
:return: a simpy process (a generator)
"""
started = env.now
# TODO: there's a lot of potential to improve fidelity here: consider image layers, simulate extraction time, etc.
# e.g., docker pull on a 13MB container takes about 5 seconds. the simulated time at 120 MBit/sec would be <1s
# find the image in the registry with the node's architecture
images = env.container_registry.find(image_str, arch=node.arch)
if not images:
raise ValueError('image not in registry: %s arch=%s' %
(image_str, node.arch))
image = images[0]
node_state = env.get_node_state(node.name)
if node_state:
if image in node_state.docker_images:
return
else:
node_state.docker_images.add(image)
size = image.size
if size <= 0:
return
# # FIXME: crude simulation of layer sharing (90% across images is shared)
# num_images = len(env.cluster.images_on_nodes[node.name]) - 1
# if num_images > 0:
# size = size * 0.1
route = env.topology.route(DockerRegistry, node)
flow = SafeFlow(env, size, route)
yield flow.start()
# for hop in route.hops:
# env.metrics.log_network(size, 'docker_pull', hop)
env.metrics.log_flow(size, env.now - started, route.source,
route.destination, 'docker_pull')
def run(self, env: Environment):
# deploy functions
deployments = self.prepare_deployments()
for deployment in deployments:
yield from env.faas.deploy(deployment)
# block until replicas become available (scheduling has finished and replicas have been deployed on the node)
logger.info('waiting for replica')
yield env.process(env.faas.poll_available_replica('python-pi'))
# generate profile
ia_generator = expovariate_arrival_profile(
constant_rps_profile(rps=20))
# run profile
yield from function_trigger(env,
deployments[0],
ia_generator,
max_requests=100)
def invoke(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
token = self.queue.request()
yield token # wait for access
# because of GIL and Threads, we can easily estimate the additional time caused by concurrent requests to the
# same Function
factor = max(1, self.scale(self.queue.count, self.queue.capacity))
try:
fet = self.deployment.sample_fet(replica.node.name)
if fet is None:
logging.error(
f"FET for node {replica.node.name} for function {self.deployment.image} was not found"
)
raise ValueError(f'{replica.node.name}')
fet = float(fet) * factor
yield env.timeout(fet)
except KeyError:
pass
self.queue.release(token)
def solve(self,
env: Environment) -> Generator[simpy.events.Event, Any, Any]:
logging.info('Calculating Pod Labels')
start = time.time()
if self.clusters is None or len(self.clusters) == 0:
# TODO caching because this may bottleneck - needs to figure out if clusters/devices have changed
self.clusters: Dict[str, Cluster] = create_clusters(env)
self.devices = get_devices(env)
self.state = State(self.devices, self.clusters)
results = []
if self.settings.parallel:
self.execute_ga_parallel(results)
else:
self.execute_ga_single_threaded(results)
for result in results:
set_reqs_for_cluster(result.instance.cluster, result.requirements,
env)
end = time.time()
logging.info("Done calculating pods")
yield env.timeout(end - start)
def function_trigger(env: Environment,
deployment: FunctionDeployment,
ia_generator,
max_requests=None):
try:
if max_requests is None:
while True:
ia = next(ia_generator)
yield env.timeout(ia)
env.process(env.faas.invoke(FunctionRequest(deployment.name)))
else:
for _ in range(max_requests):
ia = next(ia_generator)
yield env.timeout(ia)
env.process(env.faas.invoke(FunctionRequest(deployment.name)))
except simpy.Interrupt:
pass
except StopIteration:
logging.error(f'{deployment.name} gen has finished')
def deploy(self, env: Environment, replica: FunctionReplica):
yield env.timeout(0)
sched_params = {
'percentage_of_nodes_to_score': 100,
'priorities': priorities,
'predicates': predicates
}
# Set arrival profiles/workload pattern
benchmark = ConstantBenchmark('mixed', duration=200, rps=50)
# Initialize topology
storage_index = StorageIndex()
topology = urban_sensing_topology(ether_nodes, storage_index)
# Initialize environment
env = Environment()
env.simulator_factory = AIPythonHTTPSimulatorFactory(
get_raith21_function_characterizations(resource_oracle, fet_oracle))
env.metrics = Metrics(env, log=RuntimeLogger(SimulatedClock(env)))
env.topology = topology
env.faas = DefaultFaasSystem(env, scale_by_requests=True)
env.container_registry = ContainerRegistry()
env.storage_index = storage_index
env.cluster = SimulationClusterContext(env)
env.scheduler = Scheduler(env.cluster, **sched_params)
sim = Simulation(env.topology, benchmark, env=env)
result = sim.run()
dfs = {
def setup(self, env: Environment, replica: FunctionReplica):
# no setup routine
yield env.timeout(0)
def release_resources(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
env.resource_state.remove_resource(replica, 'cpu', 0.2)
yield env.timeout(0)
def claim_resources(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
# no setup time, no memory because everything is cached - only cpu usage
env.resource_state.put_resource(replica, 'cpu', 0.2)
yield env.timeout(0)
def execute(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
yield env.timeout(0.2)
def solve(self,
env: Environment) -> Generator[simpy.events.Event, Any, Any]:
yield env.timeout(0)
def invoke(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
yield env.timeout(0)
def startup(self, env: Environment, replica: FunctionReplica):
yield env.timeout(0)
def teardown(self, env: Environment, replica: FunctionReplica):
yield env.timeout(0)
def run(self, env: Environment):
yield env.timeout(0)
def claim_resources(self, env: Environment, replica: FunctionReplica,
request: FunctionRequest):
env.resource_state.put_resource(replica, 'cpu', 0.7)
env.resource_state.put_resource(replica, 'memory', 0.3)
yield env.timeout(0)