Container Service for Kubernetes:Use the MSE Ingress in Knative to implement auto scaling

Prerequisites

• Knative has been deployed in your cluster. For more information, see Deploy Knative.

• The MSE Ingress controller has been installed and authorized to access MSE. For more information, see Authorize the MSE Ingress controller to access MSE.

How it works

The throughput of an individual pod is limited. If multiple requests are routed to the same pod, the corresponding server may be overloaded. Therefore, you need to precisely control the number of concurrent requests processed by each pod. For example, in Artificial intelligence generated content (AIGC) scenarios, a single request may occupy large amounts of GPU resources. You need to strictly limit the number of concurrent requests processed by each pod. By working with the MSE Ingress, Knative can accurately scale out pods to meet traffic management requirements in scenarios where large numbers of cloud-native distributed applications are deployed.

In the following figure, the MultidimPodAutoscaler (MPA) obtains the total number of concurrent requests from the MSE Ingress, calculates the number of pods required for processing the requests, and then scales pods. This implements load-aware auto scaling. The MSE Ingress can accurately route requests to different services or versions based on routing rules and conditions.

Step 1: Deploy an MSE Ingress

1. Log on to the ACK console. In the left-side navigation pane, click Clusters.

2. On the Clusters page, find the cluster that you want to manage and click its name. In the left-side navigation pane, choose Applications > Knative.

3. On the Components tab, click Deploy Knative, select MSE for the Gateway parameter, and finish the deployment as prompted.

Step 2: Use the MSE Ingress to access Services

1. Log on to the ACK console. In the left-side navigation pane, click Clusters.

2. On the Clusters page, find the cluster that you want to manage and click its name. In the left-side navigation pane, choose Applications > Knative.

3. On the Services tab of the Knative page, set Namespace to default, click Create from Template, copy the following YAML content to the template editor, and then click Create.

   The template creates a Service named helloworld-go.

   Important

   Replace {REGION-ID} with the region where your cluster resides so that the Knative Service can pull images as expected.

   apiVersion: serving.knative.dev/v1
   kind: Service
   metadata:
     name: helloworld-go
   spec:
     template:
       metadata:
         annotations:
           autoscaling.knative.dev/class: mpa.autoscaling.knative.dev # Scale pods based on MSE metrics. Pods can be scaled to zero. 
           autoscaling.knative.dev/max-scale: '20' # Set the maximum number of pods allowed to 20. 
       spec:
         containerConcurrency: 5 # Set the maximum number of concurrent requests that each pod can process to 5. 
         containers:
         - image: registry-vpc.{REGION-ID}.aliyuncs.com/knative-sample/helloworld-go:73fbdd56 # {REGION-ID} is the region where your cluster resides, such as cn-hangzhou.
           env:
           - name: TARGET
             value: "Knative"

   If the Status column of the Service displays Created, the Service is deployed.

4. On the Services page, record the domain name and gateway IP address of the helloworld-go Service in the Default Domain and Gateway columns, respectively.

5. Run the following command to access the helloworld-go Service:

   curl -H "host: helloworld-go.default.example.com" http://8.141.XX.XX # Specify the actual gateway IP address and domain name that you obtained.

   Expected output:

   Hello Knative!

Step 3: Perform auto scaling based on the number of concurrent requests

1. Install the load testing tool hey.

   For more information about hey, see Hey.

2. Run the following command to perform a stress test on the Service:

   # Send 100,000 requests, and set the concurrency to 50 and request timeout period to 180 seconds. 
   hey -n 100000 -c 50 -t 180 -host "helloworld-go.default.example.com" "http://8.141.XX.XX"

   Expected output:

   Summary:
     Total:        86.0126 secs
     Slowest:      0.1672 secs
     Fastest:      0.0276 secs
     Average:      0.0337 secs
     Requests/sec: 1162.6199
     
     Total data:   1500000 bytes
     Size/request: 15 bytes
   
   Response time histogram:
     0.028 [1]     |
     0.042 [95291] |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
     0.056 [4573]  |■■
     0.069 [64]    |
     0.083 [19]    |
     0.097 [2]     |
     0.111 [0]     |
     0.125 [0]     |
     0.139 [18]    |
     0.153 [23]    |
     0.167 [9]     |
   
   
   Latency distribution:
     10% in 0.0294 secs
     25% in 0.0305 secs
     50% in 0.0327 secs
     75% in 0.0367 secs
     90% in 0.0386 secs
     95% in 0.0405 secs
     99% in 0.0433 secs
   
   Details (average, fastest, slowest):
     DNS+dialup:   0.0000 secs, 0.0276 secs, 0.1672 secs
     DNS-lookup:   0.0000 secs, 0.0000 secs, 0.0000 secs
     req write:    0.0000 secs, 0.0000 secs, 0.0009 secs
     resp wait:    0.0336 secs, 0.0276 secs, 0.1671 secs
     resp read:    0.0000 secs, 0.0000 secs, 0.0009 secs
   
   Status code distribution:
     [200] 100000 responses

   The output indicates that 100,000 are sent. All requests are processed.

3. Run the following command to query the scaling of pods.

   Note

   The command runs permanently until you manually terminate it. You can press Ctrl + C to terminate the command.

   kubectl get pods --watch

   

Step 4 (optional): View the monitoring data of the Service

Knative provides out-of-the-box monitoring features. On the Knative page, click the Monitoring Dashboards tab to view the monitoring data of the helloworld-go Service. For more information about how to enable the Knative dashboard, see View the Knative dashboard in Managed Service for Prometheus.