def request_generator(env: FaasSimEnvironment, arrival_profile: RandomSampler, request_factory):
sampler = BufferedSampler(arrival_profile)
try:
while True:
ia = sampler.sample()
logger.debug('next request: %.4f', ia)
env.request_queue.put(request_factory())
yield env.timeout(ia)
except simpy.Interrupt:
logger.info('request generator interrupted')
pass
def __init__(self,
max_image_variety=None,
pareto=True,
image_synthesizer=None) -> None:
super().__init__()
self.max_image_variety = max_image_variety
if pareto:
# 80% of pods will be assigned the same 20% of images
self.image_id_sampler = BufferedSampler(
IntegerTruncationSampler(
ScaledParetoSampler(0, max_image_variety)))
else:
self.image_id_sampler = None
self.image_synthesizer = image_synthesizer or ImageSynthesizer()
def __init__(self, execution_times: Dict[Tuple[str, str],
Tuple[float, float, PDist]]):
self.execution_times = execution_times
self.execution_time_samplers = {
k: BoundRejectionSampler(BufferedSampler(dist), xmin, xmax)
for k, (xmin, xmax, dist) in execution_times.items()
}
def __init__(self) -> None:
super().__init__()
self.execution_time_samplers = {
k: BoundRejectionSampler(BufferedSampler(dist), xmin, xmax)
for k, (xmin, xmax, dist) in execution_time_distributions.items()
}
def __init__(self) -> None:
self.sampler = BufferedSampler(IntegerSampler(10, 1000))
class RandomFloatDict:
def __init__(self) -> None:
self.sampler = BufferedSampler(IntegerSampler(10, 1000))
def __getitem__(self, item):
return self.sampler.sample()
class MLWorkflowPodSynthesizer:
def __init__(self,
max_image_variety=None,
pareto=True,
image_synthesizer=None) -> None:
super().__init__()
self.max_image_variety = max_image_variety
if pareto:
# 80% of pods will be assigned the same 20% of images
self.image_id_sampler = BufferedSampler(
IntegerTruncationSampler(
ScaledParetoSampler(0, max_image_variety)))
else:
self.image_id_sampler = None
self.image_synthesizer = image_synthesizer or ImageSynthesizer()
def get_image_states(self) -> Dict[str, ImageState]:
return self.image_synthesizer.get_image_states()
def create_workflow_pods(self, instance_id) -> Tuple[Pod, Pod, Pod]:
image_id = self.get_image_id(instance_id)
return (self.create_ml_wf_1_pod(instance_id,
image_id,
pod_id=instance_id * 3 + 0),
self.create_ml_wf_2_pod(instance_id,
image_id,
pod_id=instance_id * 3 + 1),
self.create_ml_wf_3_pod(instance_id,
image_id,
pod_id=instance_id * 3 + 2))
def get_image_id(self, instance_id):
if self.image_id_sampler:
image_id = self.image_id_sampler.sample()
else:
image_id = instance_id % self.max_image_variety if self.max_image_variety else instance_id
return image_id
def create_ml_wf_1_pod(self, instance_id: int, image_id: int,
pod_id: int) -> Pod:
image, _ = self.image_synthesizer.create_ml_wf_1_image(image_id)
raw_data = f'bucket_{instance_id}/raw_data'
train_data = f'bucket_{instance_id}/train_data'
return create_pod(pod_id,
image_name=image,
memory='100Mi',
labels={
'data.skippy.io/receives-from-storage': '12Mi',
'data.skippy.io/sends-to-storage': '209Mi',
'data.skippy.io/receives-from-storage/path':
raw_data,
'data.skippy.io/sends-to-storage/path':
train_data
})
def create_ml_wf_2_pod(self, instance_id: int, image_id: int,
pod_id: int) -> Pod:
image, _ = self.image_synthesizer.create_ml_wf_2_image(image_id)
train_data = f'bucket_{instance_id}/train_data'
serialized_model = f'bucket_{instance_id}/model'
return create_pod(pod_id,
image_name=image,
memory='1Gi',
labels={
'capability.skippy.io/nvidia-cuda':
'10',
'capability.skippy.io/nvidia-gpu':
'',
'data.skippy.io/receives-from-storage':
'209Mi',
'data.skippy.io/sends-to-storage':
'1500Ki',
'data.skippy.io/receives-from-storage/path':
train_data,
'data.skippy.io/sends-to-storage/path':
serialized_model
})
def create_ml_wf_3_pod(self, instance_id: int, image_id: int,
pod_id: int) -> Pod:
image, _ = self.image_synthesizer.create_ml_wf_3_image(image_id)
serialized_model = f'bucket_{instance_id}/model'
return create_pod(pod_id,
image_name=image,
labels={
'data.skippy.io/receives-from-storage':
'1500Ki',
'data.skippy.io/receives-from-storage/path':
serialized_model
})
def exp_sampler(lambd):
sampler = BufferedSampler(PDist.expon(lambd))
while True:
yield sampler.sample()