def setup_ecs_inference_service(
docker_image_uri,
framework,
cluster_arn,
model_name,
worker_instance_id,
ei_accelerator_type=None,
num_gpus=None,
region=DEFAULT_REGION,
):
"""
Function to setup Inference service on ECS
:param docker_image_uri:
:param framework:
:param cluster_arn:
:param model_name:
:param worker_instance_id:
:param num_gpus:
:param region:
:return: <tuple> service_name, task_family, revision if all steps passed else Exception
Cleans up the resources if any step fails
"""
datetime_suffix = datetime.datetime.now().strftime("%Y%m%d-%H-%M-%S")
processor = "gpu" if "gpu" in docker_image_uri else "eia" if "eia" in docker_image_uri else "cpu"
port_mappings = get_ecs_port_mappings(framework)
log_group_name = f"/ecs/{framework}-inference-{processor}"
num_cpus = ec2_utils.get_instance_num_cpus(worker_instance_id,
region=region)
# We assume that about 80% of RAM is free on the instance, since we are not directly querying it to find out
# what the memory utilization is.
memory = int(
ec2_utils.get_instance_memory(worker_instance_id, region=region) * 0.8)
cluster_name = get_ecs_cluster_name(cluster_arn, region=region)
# Below values here are just for sanity
arguments_dict = {
"family_name": cluster_name,
"image": docker_image_uri,
"log_group_name": log_group_name,
"log_stream_prefix": datetime_suffix,
"port_mappings": port_mappings,
"num_cpu": num_cpus,
"memory": memory,
"region": region
}
if processor == "gpu" and num_gpus:
arguments_dict["num_gpu"] = num_gpus
if framework == "tensorflow":
arguments_dict[
"environment"] = get_ecs_tensorflow_environment_variables(
processor, model_name)
print(f"Added environment variables: {arguments_dict['environment']}")
elif framework in ["mxnet", "pytorch"]:
arguments_dict["container_command"] = [
get_mms_run_command(model_name, processor)
]
if processor == "eia":
arguments_dict["health_check"] = {
"retries":
2,
"command": [
"CMD-SHELL",
"LD_LIBRARY_PATH=/opt/ei_health_check/lib /opt/ei_health_check/bin/health_check"
],
"timeout":
5,
"interval":
30,
"startPeriod":
60
}
arguments_dict["inference_accelerators"] = {
"deviceName": "device_1",
"deviceType": ei_accelerator_type
}
try:
task_family, revision = register_ecs_task_definition(**arguments_dict)
print(f"Created Task definition - {task_family}:{revision}")
service_name = create_ecs_service(cluster_name,
f"service-{cluster_name}",
f"{task_family}:{revision}",
region=region)
print(
f"Created ECS service - {service_name} with cloudwatch log group - {log_group_name} "
f"log stream prefix - {datetime_suffix}/{cluster_name}")
if check_running_task_for_ecs_service(cluster_name,
service_name,
region=region):
print("Service status verified as running. Running inference ...")
else:
raise Exception(f"No task running in the service: {service_name}")
return service_name, task_family, revision
except Exception as e:
raise ECSServiceCreationException(
f"Setup Inference Service Exception - {e}")
def ecs_training_test_executor(cluster_name,
cluster_arn,
training_command,
image_uri,
instance_id,
num_gpus=None):
"""
Function to run training task on ECS; Cleans up the resources after each execution
:param cluster_name:
:param cluster_arn:
:param datetime_suffix:
:param training_command:
:param image_uri:
:param instance_id:
:param num_gpus:
:return:
"""
# Set defaults to satisfy finally case
task_arn = None
task_family = None
revision = None
# Define constants for arguments to be sent to task def
image_tag = image_uri.split(':')[-1]
log_group_name = os.path.join(os.sep, 'ecs', image_tag)
datetime_suffix = datetime.datetime.now().strftime('%Y%m%d-%H-%M-%S')
num_cpus = ec2_utils.get_instance_num_cpus(instance_id)
memory = int(ec2_utils.get_instance_memory(instance_id) * 0.8)
arguments_dict = {
"family_name": cluster_name,
"image": image_uri,
"log_group_name": log_group_name,
"log_stream_prefix": datetime_suffix,
"num_cpu": num_cpus,
"memory": memory,
"entrypoint": ["sh", "-c"],
"container_command": training_command
}
if "gpu" in image_tag and num_gpus:
arguments_dict["num_gpu"] = str(num_gpus)
try:
task_family, revision = register_ecs_task_definition(**arguments_dict)
print(f"Created Task definition - {task_family}:{revision}")
task_arn = create_ecs_task(cluster_name, f"{task_family}:{revision}")
print(
f"Created ECS task - {task_arn} with cloudwatch log group - {log_group_name} log stream prefix - "
f"{os.path.join(datetime_suffix, cluster_name)}")
print("Waiting for task to stop ...")
if ecs_task_waiter(cluster_name, [task_arn], "tasks_stopped"):
ret_codes = describe_ecs_task_exit_status(cluster_name, task_arn)
if ret_codes:
# Assemble error message if we have nonzero return codes
error_msg = "Failures:\n"
for code in ret_codes:
add_on = "------------------\n"
for key, value in code.items():
add_on += f"{key}: {value}\n"
error_msg += add_on
raise ECSTrainingTestFailure(error_msg)
# Return gracefully if task stops
return
# Raise error if the task does not stop
raise ECSTaskNotStoppedError(f"Task not stopped {task_arn}")
finally:
tear_down_ecs_training_task(cluster_arn, task_arn, task_family,
revision)
