Deploying Elastic Stable Diffusion on Alibaba Cloud Serverless Kubernetes

This article explores the deployment of an enterprise-class elastic Stable Diffusion service using Alibaba Cloud's Serverless Kubernetes (ASK) and Knative. Stable Diffusion, a powerful AI model for image generation, combined with the elasticity of serverless architecture, offers a scalable and efficient solution for businesses. By leveraging ASK and Knative, we can create a highly responsive and cost-effective AI image generation service that automatically scales based on demand.

To set up the environment, we need to complete several crucial steps: 1. Create an MSE Cloud-Native Gateway: This involves logging into the Microservices Engine (MSE) console and activating MSE cloud native gateways. 2. Authorize MSE to Access ECI: This step ensures that MSE can interact with Elastic Container Instance (ECI). 3. Create an ASK Pro Cluster: We'll use the ACK console to create a Serverless Kubernetes cluster, selecting the Professional edition for enhanced features. Key configurations include setting the cluster name, region (recommended: China Hong Kong), and enabling Knative. These preparatory steps lay the foundation for our elastic Stable Diffusion service, ensuring we have the necessary infrastructure and permissions in place.

Deploying the Stable Diffusion service involves creating a Knative Service within our ASK cluster. Here's how to do it: 1. Navigate to the Knative page in the cluster's dashboard. 2. Create a new Service from a template, using a provided YAML configuration. 3. The YAML defines a service named 'knative-sd-demo' with specific annotations for affinity and autoscaling. 4. The container image used is a pre-configured Stable Diffusion setup. 5. After creation, wait for the service status to change to 'Created'. This deployment leverages Knative's serverless capabilities, allowing the Stable Diffusion service to scale automatically based on demand.

To access and test the Stable Diffusion service, we need to create a traffic generator and expose it: 1. Deploy a 'portal-server' using a provided YAML configuration. This serves as both a load generator and a web interface for the Stable Diffusion service. 2. The deployment creates a LoadBalancer service, providing an external IP address. 3. Access the web interface using the LoadBalancer's IP address. 4. Use the interface to generate images by providing prompts to the Stable Diffusion model. This setup allows us to interact with the Stable Diffusion service and prepare for stress testing.

To verify the elasticity of our ASK cluster and the Stable Diffusion service, we'll perform stress tests: 1. Use the stress test feature in the web interface. 2. Set concurrency to 5 and total number of requests to 20. 3. Initiate the stress test and observe the results. 4. During the test, monitor the creation of new pods to handle increased load. 5. Observe the generated images displayed on the page. This stress test demonstrates the ASK cluster's ability to automatically scale the Stable Diffusion service in response to increased demand, showcasing the power of serverless architecture for AI workloads.

After completing the experiment, it's important to manage resources properly to avoid unexpected costs: 1. You can choose to keep the ASK Pro cluster for further use, understanding the associated costs. 2. If you decide to release resources, follow these steps: - Delete the ASK Pro cluster from the Clusters page. - Release the MSE Cloud Native Gateway through the MSE console. - Remove the NAT gateway using the NAT Gateway console. Proper resource management ensures cost efficiency and helps maintain a clean cloud environment after experimentation.