def get_tf_gpu_deployment(fet_oracle: FetOracle, resource_oracle: ResourceOracle,
ranking: DeploymentRanking):
tf_gpu_characterization = FunctionCharacterization(images.tf_gpu_manifest,
fet_oracle,
resource_oracle)
tf_gpu_function = FunctionDefinition(
name=images.tf_gpu_function,
image=images.tf_gpu_manifest,
characterization=tf_gpu_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '3', 'device.edgerun.io/accelerator': 'GPU',
'device.edgerun.io/vram': '2000', })
tf_gpu_function.requests = Resources.from_str("300Mi", "1000m")
deployment = FunctionDeployment(
images.tf_gpu_function,
{
images.tf_gpu_manifest: tf_gpu_function,
},
ranking
)
deployment.function_factor = {
images.tf_gpu_manifest: 1
}
return deployment
def get_fio_deployment(fet_oracle: FetOracle, resource_oracle: ResourceOracle,
ranking: DeploymentRanking) -> FunctionDeployment:
fio_characterization = FunctionCharacterization(images.fio_manifest,
fet_oracle,
resource_oracle)
fio_function = FunctionDefinition(
name=images.fio_function,
image=images.fio_manifest,
characterization=fio_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '1'})
fio_function.requests = Resources.from_str("200Mi", "1000m")
deployment = FunctionDeployment(
images.fio_function,
{
images.fio_manifest: fio_function,
},
ranking
)
deployment.function_factor = {
images.fio_manifest: 1
}
return deployment
def get_speech_inference_deployment(fet_oracle: FetOracle, resource_oracle: ResourceOracle,
ranking: DeploymentRanking) -> FunctionDeployment:
speech_inference_gpu_characterization = FunctionCharacterization(images.speech_inference_gpu_manifest, fet_oracle,
resource_oracle)
speech_inference_tflite_characterization = FunctionCharacterization(images.speech_inference_tflite_manifest,
fet_oracle, resource_oracle)
tflite = storage.speech_model_tflite_bucket_item
data_storage_tflite = {
'data.skippy.io/receives-from-storage': '48M',
'data.skippy.io/receives-from-storage/path': f'{storage.speech_bucket}/{tflite.name}',
}
gpu = storage.speech_model_gpu_bucket_item
data_storage_gpu = {
# this size is without scorer object, which is used to impove accuracy but doesn't seem to affect runtime,
# scorer weighs around 900M - simple benchmarks in bash have made no difference in runtime
'data.skippy.io/receives-from-storage': '188M',
'data.skippy.io/receives-from-storage/path': f'{storage.speech_bucket}/{gpu.name}',
}
speech_gpu_function = FunctionDefinition(
name=images.speech_inference_function,
image=images.speech_inference_gpu_manifest,
characterization=speech_inference_gpu_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '0', 'device.edgerun.io/accelerator': 'GPU',
'device.edgerun.io/vram': '1500', })
speech_gpu_function.labels.update(data_storage_gpu)
speech_gpu_function.requests = Resources.from_str("300Mi", "1000m")
speech_tflite_function = FunctionDefinition(
name=images.speech_inference_function,
image=images.speech_inference_tflite_manifest,
characterization=speech_inference_tflite_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '1'})
speech_tflite_function.labels.update(data_storage_tflite)
speech_tflite_function.requests = Resources.from_str("100Mi", "1000m")
deployment = FunctionDeployment(
images.speech_inference_function,
{
images.speech_inference_gpu_manifest: speech_gpu_function,
images.speech_inference_tflite_manifest: speech_tflite_function,
},
ranking
)
deployment.function_factor = {
images.speech_inference_tflite_manifest: 1,
images.speech_inference_gpu_manifest: 1
}
return deployment
def get_mobilenet_inference_deployment(fet_oracle: FetOracle, resource_oracle: ResourceOracle,
ranking: DeploymentRanking) -> FunctionDeployment:
mobilenet_inference_tflite_characterization = FunctionCharacterization(images.mobilenet_inference_tflite_manifest,
fet_oracle,
resource_oracle)
mobilenet_inference_tpu_characterization = FunctionCharacterization(images.mobilenet_inference_tpu_manifest,
fet_oracle, resource_oracle)
tflite = storage.mobilenet_model_tflite_bucket_item.name
data_storage_tflite_labels = {
'data.skippy.io/receives-from-storage': '4M',
'data.skippy.io/receives-from-storage/path': f'{storage.mobilenet_bucket}/{tflite}',
}
tpu = storage.mobilenet_model_tpu_bucket_item.name
data_storage_tpu_labels = {
'data.skippy.io/receives-from-storage': '4M',
'data.skippy.io/receives-from-storage/path': f'{storage.mobilenet_bucket}/{tpu}',
}
mobilenet_tpu_function = FunctionDefinition(
name=images.mobilenet_inference_function,
image=images.mobilenet_inference_tpu_manifest,
characterization=mobilenet_inference_tpu_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '1b', 'device.edgerun.io/accelerator': 'TPU'})
mobilenet_tpu_function.requests = Resources.from_str("100Mi", "1000m")
mobilenet_tflite_function = FunctionDefinition(
name=images.mobilenet_inference_function,
image=images.mobilenet_inference_tflite_manifest,
characterization=mobilenet_inference_tflite_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '1a'})
mobilenet_tflite_function.requests = Resources.from_str("100Mi", "1000m")
mobilenet_tpu_function.labels.update(data_storage_tpu_labels)
mobilenet_tflite_function.labels.update(data_storage_tflite_labels)
deployment = FunctionDeployment(
images.mobilenet_inference_function,
{
images.mobilenet_inference_tpu_manifest: mobilenet_tpu_function,
images.mobilenet_inference_tflite_manifest: mobilenet_tflite_function,
},
ranking
)
deployment.function_factor = {
images.mobilenet_inference_tpu_manifest: 1,
images.mobilenet_inference_tflite_manifest: 1
}
return deployment
def get_resnet_training_deployment(fet_oracle: FetOracle, resource_oracle: ResourceOracle,
ranking: DeploymentRanking) -> FunctionDeployment:
resnet_training_gpu_characterization = FunctionCharacterization(images.resnet50_training_gpu_manifest,
fet_oracle,
resource_oracle)
resnet_training_cpu_characterization = FunctionCharacterization(images.resnet50_training_cpu_manifest,
fet_oracle,
resource_oracle)
data = storage.resnet_train_bucket_item.name
data_storage_labels = {
'data.skippy.io/receives-from-storage': '58M',
'data.skippy.io/sends-to-storage': '103M',
'data.skippy.io/receives-from-storage/path': f'{storage.resnet_train_bucket}/{data}',
'data.skippy.io/sends-to-storage/path': f'{storage.resnet_train_bucket}/updated_model'
}
resnet_training_gpu_function = FunctionDefinition(
name=images.resnet50_training_function,
image=images.resnet50_training_gpu_manifest,
characterization=resnet_training_gpu_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '2', 'device.edgerun.io/accelerator': 'GPU',
'device.edgerun.io/vram': '2000', })
resnet_training_gpu_function.labels.update(data_storage_labels)
resnet_training_gpu_function.requests = Resources.from_str("800Mi", "1000m")
resnet_training_cpu_function = FunctionDefinition(
name=images.resnet50_training_function,
image=images.resnet50_training_cpu_manifest,
characterization=resnet_training_cpu_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '2'})
resnet_training_cpu_function.labels.update(data_storage_labels)
resnet_training_cpu_function.requests = Resources.from_str("1Gi", "1000m")
deployment = FunctionDeployment(
images.resnet50_training_function,
{
images.resnet50_training_gpu_manifest: resnet_training_gpu_function,
images.resnet50_training_cpu_manifest: resnet_training_cpu_function
},
ranking
)
deployment.function_factor = {
images.resnet50_training_gpu_manifest: 1,
images.resnet50_training_cpu_manifest: 1
}
return deployment
def get_tf_gpu_deployment(ranking: DeploymentRanking,
scaling_config: ScalingConfiguration = None):
# Design Time
tf_gpu_function_image = FunctionImage(image=images.tf_gpu_manifest)
tf_gpu_function = Function(name=images.tf_gpu_function,
fn_images=[tf_gpu_function_image])
# Run time
tf_gpu_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="300Mi")
tf_gpu_function_container = FunctionContainer(
tf_gpu_function_image,
resource_config=tf_gpu_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '3',
'device.edgerun.io/accelerator': 'GPU',
'device.edgerun.io/vram': '2000',
})
deployment = FunctionDeployment(
tf_gpu_function, [tf_gpu_function_container],
ScalingConfiguration() if scaling_config is None else scaling_config,
ranking)
return deployment
def prepare_deployments(self) -> List[FunctionDeployment]:
python_pi_deployment = FunctionDeployment(
name='python-pi',
function_definitions={
'python-pi-cpu': FunctionDefinition(name='python-pi', image='python-pi-cpu')
}
)
resnet50_inference_deployment = FunctionDeployment(
name='resnet50-inference',
function_definitions={
'resnet50-inference-gpu': FunctionDefinition(name='resnet50-inference', image='resnet50-inference-gpu'),
'resnet50-inference-cpu': FunctionDefinition(name='resnet50-inference', image='resnet50-inference-cpu')
}
)
return [python_pi_deployment, resnet50_inference_deployment]
def get_fio_deployment(
ranking: DeploymentRanking,
scaling_config: ScalingConfiguration = None) -> FunctionDeployment:
# Design Time
fio_function_image = FunctionImage(image=images.fio_manifest)
fio_function = Function(name=images.fio_function,
fn_images=[fio_function_image])
# Run time
fio_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="200Mi")
fio_function_container = FunctionContainer(
fio_function_image,
resource_config=fio_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '1'
})
deployment = FunctionDeployment(
fio_function, [fio_function_container],
ScalingConfiguration() if scaling_config is None else scaling_config,
ranking)
return deployment
def prepare_resnet_inference_deployment(self):
# Design time
resnet_inference = 'resnet50-inference'
inference_cpu = 'resnet50-inference-cpu'
inference_gpu = 'resnet50-inference-gpu'
resnet_inference_gpu = FunctionImage(image=inference_gpu)
resnet_inference_cpu = FunctionImage(image=inference_cpu)
resnet_fn = Function(
resnet_inference,
fn_images=[resnet_inference_gpu, resnet_inference_cpu])
# Run time
# default kubernetes requested resources
resnet_cpu_container = FunctionContainer(resnet_inference_cpu)
# custom defined requested resources
request = KubernetesResourceConfiguration.create_from_str(
cpu='100m', memory='1024Mi')
resnet_gpu_container = FunctionContainer(resnet_inference_gpu,
resource_config=request)
resnet_fd = FunctionDeployment(
resnet_fn, [resnet_cpu_container, resnet_gpu_container],
ScalingConfiguration(),
DeploymentRanking([inference_gpu, inference_cpu]))
return resnet_fd
def get_resnet50_inference_deployment(fet_oracle: FetOracle, resource_oracle: ResourceOracle,
ranking: DeploymentRanking) -> FunctionDeployment:
resnet_cpu_characterization = FunctionCharacterization(images.resnet50_inference_cpu_manifest, fet_oracle,
resource_oracle)
resnet_gpu_characterization = FunctionCharacterization(images.resnet50_inference_gpu_manifest, fet_oracle,
resource_oracle)
model = storage.resnet_model_bucket_item.name
data_storage = {
'data.skippy.io/receives-from-storage': '103M',
'data.skippy.io/receives-from-storage/path': f'{storage.resnet_model_bucket}/{model}',
}
resnet50_cpu_function = FunctionDefinition(
name=images.resnet50_inference_function,
image=images.resnet50_inference_cpu_manifest,
characterization=resnet_cpu_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '4a'})
resnet50_cpu_function.requests = Resources.from_str("150Mi", "1000m")
resnet50_gpu_function = FunctionDefinition(
name=images.resnet50_inference_function,
image=images.resnet50_inference_gpu_manifest,
characterization=resnet_gpu_characterization,
labels={'watchdog': 'http', 'workers': '4', 'device.edgerun.io/accelerator': 'GPU',
'device.edgerun.io/vram': '1500',
'cluster': '4b'})
resnet50_gpu_function.requests = Resources.from_str("400Mi", "1000m")
resnet50_gpu_function.labels.update(data_storage)
resnet50_cpu_function.labels.update(data_storage)
deployment = FunctionDeployment(
images.resnet50_inference_function,
{
images.resnet50_inference_gpu_manifest: resnet50_gpu_function,
images.resnet50_inference_cpu_manifest: resnet50_cpu_function,
},
ranking
)
deployment.function_factor = {
images.resnet50_inference_gpu_manifest: 1,
images.resnet50_inference_cpu_manifest: 1,
}
return deployment
def get_resnet50_inference_deployment(
ranking: DeploymentRanking,
scaling_config: ScalingConfiguration = None) -> FunctionDeployment:
# Design Time
resnet50_cpu_function_image = FunctionImage(
image=images.resnet50_inference_cpu_manifest)
resnet50_gpu_function_image = FunctionImage(
image=images.resnet50_inference_gpu_manifest)
resnet50_function = Function(
images.resnet50_inference_function,
fn_images=[resnet50_gpu_function_image, resnet50_cpu_function_image])
# Run time
model = storage.resnet_model_bucket_item.name
data_storage = {
'data.skippy.io/receives-from-storage':
'103M',
'data.skippy.io/receives-from-storage/path':
f'{storage.resnet_model_bucket}/{model}',
}
resnet50_cpu_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="150Mi")
resnet50_cpu_function = FunctionContainer(
resnet50_cpu_function_image,
resource_config=resnet50_cpu_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '4a'
})
resnet50_gpu_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="400Mi")
resnet50_gpu_function = FunctionContainer(
resnet50_gpu_function_image,
resource_config=resnet50_gpu_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'device.edgerun.io/accelerator': 'GPU',
'device.edgerun.io/vram': '1500',
'cluster': '4b'
})
resnet50_gpu_function.labels.update(data_storage)
resnet50_cpu_function.labels.update(data_storage)
deployment = FunctionDeployment(
resnet50_function, [resnet50_gpu_function, resnet50_cpu_function],
ScalingConfiguration() if scaling_config is None else scaling_config,
ranking)
return deployment
def get_resnet_preprocessing_deployment(fet_oracle: FetOracle, resource_oracle: ResourceOracle,
ranking: DeploymentRanking):
resnet_preprocessing_characterization = FunctionCharacterization(images.resnet50_preprocessing_manifest,
fet_oracle,
resource_oracle)
data = storage.resnet_pre_bucket_item.name
data_storage_labels = {
'data.skippy.io/receives-from-storage': '14M',
'data.skippy.io/sends-to-storage': '14M',
'data.skippy.io/receives-from-storage/path': f'{storage.resnet_pre_bucket}/{data}',
'data.skippy.io/sends-to-storage/path': f'{storage.resnet_pre_bucket}/preprocessed'
}
resnet_preprocessing_function = FunctionDefinition(
name=images.resnet50_preprocessing_function,
image=images.resnet50_preprocessing_manifest,
characterization=resnet_preprocessing_characterization,
labels={'watchdog': 'http', 'workers': '4', 'cluster': '1'}
)
resnet_preprocessing_function.labels.update(data_storage_labels)
resnet_preprocessing_function.requests = Resources.from_str("100Mi", "1000m")
deployment = FunctionDeployment(
images.resnet50_preprocessing_function,
{
images.resnet50_preprocessing_manifest: resnet_preprocessing_function,
},
ranking
)
deployment.function_factor = {
images.resnet50_preprocessing_manifest: 1
}
return deployment
def prepare_python_pi_deployment(self):
# Design Time
python_pi = 'python-pi'
python_pi_cpu = FunctionImage(image='python-pi-cpu')
python_pi_fn = Function(python_pi, fn_images=[python_pi_cpu])
# Run time
python_pi_fn_container = FunctionContainer(python_pi_cpu)
python_pi_fd = FunctionDeployment(python_pi_fn,
[python_pi_fn_container],
ScalingConfiguration())
return python_pi_fd
def get_resnet_preprocessing_deployment(
ranking: DeploymentRanking,
scaling_config: ScalingConfiguration = None):
# Design time
resnet_preprocessing_function_image = FunctionImage(
image=images.resnet50_preprocessing_manifest)
resnet_preprocessing_function = Function(
name=images.resnet50_preprocessing_function,
fn_images=[resnet_preprocessing_function_image])
# Run time
data = storage.resnet_pre_bucket_item.name
data_storage_labels = {
'data.skippy.io/receives-from-storage':
'14M',
'data.skippy.io/sends-to-storage':
'14M',
'data.skippy.io/receives-from-storage/path':
f'{storage.resnet_pre_bucket}/{data}',
'data.skippy.io/sends-to-storage/path':
f'{storage.resnet_pre_bucket}/preprocessed'
}
resnet_preprocessing_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="100Mi")
resnet_preprocessing_function_container = FunctionContainer(
fn_image=resnet_preprocessing_function_image,
resource_config=resnet_preprocessing_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '1'
})
resnet_preprocessing_function.labels.update(data_storage_labels)
deployment = FunctionDeployment(
resnet_preprocessing_function,
[resnet_preprocessing_function_container],
ScalingConfiguration() if scaling_config is None else scaling_config,
ranking)
return deployment
def prepare_resnet_training_deployment(self):
# Design time
resnet_training = 'resnet50-training'
training_cpu = 'resnet50-training-cpu'
resnet_training_cpu = FunctionImage(image=training_cpu)
resnet_fn = Function(resnet_training, fn_images=[resnet_training_cpu])
# Run time
resnet_cpu_container = FunctionContainer(resnet_training_cpu)
resnet_fd = FunctionDeployment(resnet_fn, [resnet_cpu_container],
ScalingConfiguration(),
DeploymentRanking([training_cpu]))
return resnet_fd
def prepare_resnet_inference_deployment(self):
# Design time
resnet_inference = 'resnet50-inference'
inference_cpu = 'resnet50-inference-cpu'
resnet_inference_cpu = FunctionImage(image=inference_cpu)
resnet_fn = Function(resnet_inference,
fn_images=[resnet_inference_cpu])
# Run time
resnet_cpu_container = FunctionContainer(resnet_inference_cpu)
resnet_fd = FunctionDeployment(resnet_fn, [resnet_cpu_container],
ScalingConfiguration(),
DeploymentRanking([inference_cpu]))
return resnet_fd
def get_speech_inference_deployment(
ranking: DeploymentRanking,
scaling_config: ScalingConfiguration = None) -> FunctionDeployment:
# Design time
speech_gpu_function_image = FunctionImage(
image=images.speech_inference_gpu_manifest)
speech_tflite_function_image = FunctionImage(
image=images.speech_inference_tflite_manifest)
speech_function = Function(
images.speech_inference_function,
fn_images=[speech_gpu_function_image, speech_tflite_function_image])
# Run time
tflite = storage.speech_model_tflite_bucket_item
data_storage_tflite = {
'data.skippy.io/receives-from-storage':
'48M',
'data.skippy.io/receives-from-storage/path':
f'{storage.speech_bucket}/{tflite.name}',
}
gpu = storage.speech_model_gpu_bucket_item
data_storage_gpu = {
# this size is without scorer object, which is used to impove accuracy but doesn't seem to affect runtime,
# scorer weighs around 900M - simple benchmarks in bash have made no difference in runtime
'data.skippy.io/receives-from-storage':
'188M',
'data.skippy.io/receives-from-storage/path':
f'{storage.speech_bucket}/{gpu.name}',
}
speech_gpu_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="300Mi")
speech_gpu_function = FunctionContainer(
speech_gpu_function_image,
resource_config=speech_gpu_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '0',
'device.edgerun.io/accelerator': 'GPU',
'device.edgerun.io/vram': '1500',
})
speech_gpu_function.labels.update(data_storage_gpu)
speech_tflite_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="100Mi")
speech_tflite_function = FunctionContainer(
speech_tflite_function_image,
resource_config=speech_tflite_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '1'
})
speech_tflite_function.labels.update(data_storage_tflite)
deployment = FunctionDeployment(
speech_function, [speech_gpu_function, speech_tflite_function],
ScalingConfiguration() if scaling_config is None else scaling_config,
ranking)
return deployment
def get_mobilenet_inference_deployment(
ranking: DeploymentRanking,
scaling_config: ScalingConfiguration = None) -> FunctionDeployment:
# Design time
mobilenet_tpu_function_image = FunctionImage(
image=images.mobilenet_inference_tpu_manifest)
mobilenet_tflite_function_image = FunctionImage(
image=images.mobilenet_inference_tflite_manifest)
mobilenet_function = Function(images.mobilenet_inference_function,
fn_images=[
mobilenet_tpu_function_image,
mobilenet_tflite_function_image
])
# Run time
tflite = storage.mobilenet_model_tflite_bucket_item.name
data_storage_tflite_labels = {
'data.skippy.io/receives-from-storage':
'4M',
'data.skippy.io/receives-from-storage/path':
f'{storage.mobilenet_bucket}/{tflite}',
}
tpu = storage.mobilenet_model_tpu_bucket_item.name
data_storage_tpu_labels = {
'data.skippy.io/receives-from-storage':
'4M',
'data.skippy.io/receives-from-storage/path':
f'{storage.mobilenet_bucket}/{tpu}',
}
mobilenet_tpu_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="100Mi")
mobilenet_tpu_function = FunctionContainer(
mobilenet_tpu_function_image,
resource_config=mobilenet_tpu_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '1b',
'device.edgerun.io/accelerator': 'TPU'
})
mobilenet_tflite_function_requests = KubernetesResourceConfiguration.create_from_str(
cpu="1000m", memory="100Mi")
mobilenet_tflite_function = FunctionContainer(
mobilenet_tflite_function_image,
resource_config=mobilenet_tflite_function_requests,
labels={
'watchdog': 'http',
'workers': '4',
'cluster': '1a'
})
mobilenet_tpu_function.labels.update(data_storage_tpu_labels)
mobilenet_tflite_function.labels.update(data_storage_tflite_labels)
deployment = FunctionDeployment(
mobilenet_function,
[mobilenet_tpu_function, mobilenet_tflite_function],
ScalingConfiguration() if scaling_config is None else scaling_config,
ranking)
deployment.function_factor = {
images.mobilenet_inference_tpu_manifest: 1,
images.mobilenet_inference_tflite_manifest: 1
}
return deployment