SageMaker Inference Recommender - CloudWatch Metrics

This notebook's CI test result for us-west-2 is as follows. CI test results in other regions can be found at the end of the notebook.

1. Introduction

SageMaker Inference Recommender is a new capability of SageMaker that reduces the time required to get machine learning (ML) models in production by automating load tests and optimizing model performance across instance types. You can use Inference Recommender to select a real-time inference endpoint that delivers the best performance at the lowest cost.

Get started with Inference Recommender on SageMaker in minutes while selecting an instance and get an optimized endpoint configuration in hours, eliminating weeks of manual testing and tuning time.

In this notebook, we will walk you through getting recommendations for a TensorFlow model and using information from Inference Recommender to pull additional metrics from CloudWatch.

2. Setup

Note that we are using the conda_tensorflow2_p36 kernel in SageMaker Notebook Instances. This is running Python 3.6 and TensorFlow 2.1.3. If you'd like to use the same setup, in the AWS Management Console, go to the Amazon SageMaker console. Choose Notebook Instances, and click create a new notebook instance. Upload the current notebook and set the kernel. You can also run this in SageMaker Studio Notebooks with the TensorFlow 2.1 Python 3.6 CPU Optimized kernel.

In the next steps, you'll import standard methods and libraries as well as set variables that will be used in this notebook. The get_execution_role function retrieves the AWS Identity and Access Management (IAM) role you created at the time of creating your notebook instance.

3. Machine learning model details

Inference Recommender uses metadata about your ML model to recommend the best instance types and endpoint configurations for deployment. You can provide as much or as little information as you'd like but the more information you provide, the better your recommendations will be.

ML Frameworks: TENSORFLOW, PYTORCH, XGBOOST, SAGEMAKER-SCIKIT-LEARN

ML Domains: COMPUTER_VISION, NATURAL_LANGUAGE_PROCESSING, MACHINE_LEARNING

Example ML Tasks: CLASSIFICATION, REGRESSION, IMAGE_CLASSIFICATION, OBJECT_DETECTION, SEGMENTATION, MASK_FILL, TEXT_CLASSIFICATION, TEXT_GENERATION, OTHER

Note: Select the task that is the closest match to your model. Chose OTHER if none apply.

4. Create a model archive

SageMaker models need to be packaged in .tar.gz files. When your SageMaker Endpoint is provisioned, the files in the archive will be extracted and put in /opt/ml/model/ on the Endpoint.

In this step, there are two optional tasks to:

(1) Download a pretrained model from Keras applications

(2) Download a sample inference script (inference.py) from S3

These tasks are provided as a sample reference but can and should be modified when using your own trained models with Inference Recommender.

Optional: Download model from Keras applications

Let's download the model from Keras applications. By setting the variable download_the_model=False, you can skip the download and provide your own model archive.

Create a tarball

To bring your own TensorFlow model, SageMaker expects a single archive file in .tar.gz format, containing a model file (*.pb) in TF SavedModel format and the script (*.py) for inference.

Upload to S3

We now have a model archive ready. We need to upload it to S3 before we can use with Inference Recommender. Furthermore, we will use the SageMaker Python SDK to handle the upload.

5. Create a sample payload archive

We need to create an archive that contains individual files that Inference Recommender can send to your Endpoint. Inference Recommender will randomly sample files from this archive so make sure it contains a similar distribution of payloads you'd expect in production. Note that your inference code must be able to read in the file formats from the sample payload.

Here we are only adding four images for the example. For your own use case(s), it's recommended to add a variety of samples that is representative of your payloads.

Create a tarball

Upload to S3

Next, we'll upload the packaged payload examples (payload.tar.gz) that was created above to S3. The S3 location will be used as input to our Inference Recommender job later in this notebook.

6. Register model in Model Registry

In order to use Inference Recommender, you must have a versioned model in SageMaker Model Registry. To register a model in the Model Registry, you must have a model artifact packaged in a tarball and an inference container image. Registering a model includes the following steps:

1. Create Model Group: This is a one-time task per machine learning use case. A Model Group contains one or more versions of your packaged model.

2. Register Model Version/Package: This task is performed for each new packaged model version.

Container image URL

If you don’t have an inference container image, you can use one of the open source AWS Deep Learning Containers (DLCs) provided by AWS to serve your ML model. The code below retrieves a DLC based on your ML framework, framework version, python version, and instance type.

Create Model Group

Register Model Version/Package

In this step, you'll register your pretrained model that was packaged in the prior steps as a new version in SageMaker Model Registry. First, you'll configure the model package/version identifying which model package group this new model should be registered within as well as identify the initial approval status. You'll also identify the domain and task for your model. These values were set earlier in the notebook where ml_domain = 'COMPUTER_VISION' and ml_task = 'IMAGE_CLASSIFICATION'

Note: ModelApprovalStatus is a configuration parameter that can be used in conjunction with SageMaker Projects to trigger automated deployment pipeline.

Set up inference specification

You'll now setup the inference specification configuration for your model version. This contains information on how the model should be hosted.

Inference Recommender expects a single input MIME type for sending requests. Learn more about common inference data formats on SageMaker. This MIME type will be sent in the Content-Type header when invoking your endpoint.

If you specify a set of instance types below (i.e. non-empty list), then Inference Recommender will only support recommendations within the set of instances below. For this example, we provide a list of common instance types used for image classification algorithms.

Now that you've configured the model package, the next step is to create the model package/version in SageMaker Model Registry

Alternative Option: ContainerConfig

If you are missing mandatory fields to create an inference recommender job in your model package version like so (this create_model_package_input_dict is missing Domain, Task, and SamplePayloadUrl):

create_model_package_input_dict = {
    "ModelPackageGroupName": model_package_group_name,
    "ModelPackageDescription": model_package_description,
    "ModelApprovalStatus": model_approval_status,
}

You may define the fields Domain, Task, and SamplePayloadUrl in the optional field ContainerConfig like so:

payload_config = {
    "SamplePayloadUrl": sample_payload_url,
}

container_config = {
    "Domain": ml_domain.upper(),
    "Task": ml_task.upper(),
    "PayloadConfig": payload_config,
}

And then provide it directly within create_inference_recommendations_job() API like so:

default_response = client.create_inference_recommendations_job(
    JobName=str(default_job),
    JobDescription="",
    JobType="Default",
    RoleArn=role,
    InputConfig={
        "ModelPackageVersionArn": model_package_arn,
        "ContainerConfig": container_config
    },
)

For more information on what else can be provided via ContainerConfig please refer to the CreateInferenceRecommendationsJob doc here: CreateInferenceRecommendationsJob

7. Create an Inference Recommender Default Job

Now with your model in Model Registry, you can kick off a 'Default' job to get instance recommendations. This only requires your ModelPackageVersionArn and comes back with recommendations within 45 minutes.

The output is a list of instance type recommendations with associated environment variables, cost, throughput and latency metrics.

8. Instance Recommendation Results

Each inference recommendation includes InstanceType, InitialInstanceCount, EnvironmentParameters which are tuned environment variable parameters for better performance. We also include performance and cost metrics such as MaxInvocations, ModelLatency, CostPerHour and CostPerInference. We believe these metrics will help you narrow down to a specific endpoint configuration that suits your use case.

Example:

If your motivation is overall price-performance with an emphasis on throughput, then you should focus on CostPerInference metrics
If your motivation is a balance between latency and throughput, then you should focus on ModelLatency / MaxInvocations metrics

Optional: ListInferenceRecommendationsJobSteps

To see the list of subtasks for an Inference Recommender job, simply provide the JobName to the ListInferenceRecommendationsJobSteps API.

To see more information for the API, please refer to the doc here: ListInferenceRecommendationsJobSteps

9. Custom Load Test

With an 'Advanced' job, you can provide your production requirements, select instance types, tune environment variables and perform more extensive load tests. This typically takes 2 hours depending on your traffic pattern and number of instance types.

The output is a list of endpoint configuration recommendations (instance type, instance count, environment variables) with associated cost, throughput and latency metrics.

In the below example, we are tuning the endpoint against an environment variable OMP_NUM_THREADS with three values [2, 4, 10] and we aim to have a 95th percentile latency to be no more than 100 ms.

Instance configurations

Stopping conditions

Traffic pattern

Resource limits

10. Custom Load Test Results

Inference Recommender benchmarks both endpoint configurations. Analyzing the load test result we find,

OMP_NUM_THREADS = 2 shows ~20% better throughput when compared to OMP_NUM_THREADS = 4
OMP_NUM_THREADS = 2 shows 25% saving in inference-cost when compared to OMP_NUM_THREADS = 4

Overall, OMP_NUM_THREADS = 2 is much better endpoint configuration than OMP_NUM_THREADS = 4

11. Review detailed CloudWatch metrics

Below we've provided a helper script cloudwatch.py that you can import to get access to functions for reviewing CW metrics. Just by providing your region and Inference Recommender job name you can get detailed charts of your endpoints during the load test.

Learn more about CloudWatch: https://docs.aws.amazon.com/sagemaker/latest/dg/monitoring-cloudwatch.html

Below is an example of the CloudWatch metrics and graphs returned for a single endpoint benchmarked by Inference Recommender.

11b. View Inference Recommender CloudWatch metrics

You can also observe load test metrics as seen by Inference Recommender. Inference Recommender publishes the following metrics to the /aws/sagemaker/InferenceRecommendationsJobs CloudWatch name space.

We will graph the "ClientInvocations" metric of the first recommendation in the job result as an example.

9. Conclusion

This notebook discussed how to use SageMaker Inference Recommender with a TensorFlow model to help determine the right CPU instance type, instance count and environment variables to reduce costs and maximize performance. The notebook walked you through downloading a pre-trained TensorFlow model, registering your model in Model Registry, creating an Inference Recommender Default job, and creating an Inference Recommender Advanced job to get recommendations. In addition, you can use the supplied helper functions to review CloudWatch metrics right from your notebook. You can modify the environment variables and instance types to match your own ML workload as well as bring your own TensorFlow model for testing.

Notebook CI Test Results

This notebook was tested in multiple regions. The test results are as follows, except for us-west-2 which is shown at the top of the notebook.