Cloudflare Argo Load Test Instruction

Run with local docker-compose

REQUIRED local port:

3000  (grafana)
9090  (prometheus)
10001 (envoy)

Make sure the above ports are free on your machine before running docker-compose up. Or you can change these ports from the docker-compose.yml.

REQUIRED file:

~/.cloudflared/cert.pem

Make sure you have the ~/.cloudflared/cert.pem on your machine before running docker-compose up, which can be generated by running cloudflared tunnel login.

Then run the following:

export TUNNEL_HOSTNAME=tunnel.dcard.tw
docker-compose up

Test in another shell:

export TUNNEL_HOSTNAME=tunnel.dcard.tw
curl -i https://$TUNNEL_HOSTNAME/healthz # should return http 200

Let the client to request the tunnel by changing the value of Hosts field in the ./client/config.json from

https://envoy

https://tunnel.dcard.tw

Finally check the dashboard on http://localhost:3000. The username and password are both admin

Run on GKE

Use gke.tf to create the GKE cluster:

terraform init

export TF_VAR_gcp_project_id=your-gcp-project-id
terraform apply

Then, get the GKE credentials from gcloud CLI:

gcloud container clusters get-credentials argo-test --zone asia-east1-a --project your-gcp-project-id

Then, upload ~/.cloudflared/cert.pem to GKE

kubectl create secret generic tunnel-crt --from-file="$HOME/.cloudflared/cert.pem"

Then, use kube.yaml with {TUNNEL_DOMAIN} substitution to deploy pods:

cat kube.yaml | sed 's/{TUNNEL_DOMAIN}/.dcard.tw/g' | kubectl apply -f -

Then, wait all pods to be running. it can take a while for GKE node pools to scale up.

Check pod status by using

kubectl get po

When all pods are running, port-forward the grafana to check the dashboard:

kubectl port-forward grafana-0 3000

The username and password are both admin.

Then, you can adjust the replicas of client, backend, or tunnel:

kubectl scale deploy backend --replicas=3
kubectl scale deploy client --replicas=9
kubectl scale deploy tunnel --replicas=2

You can also test against GCP Load Balancer directly by changing the Hosts field of client ConfigMap in kube.yaml to the IP address of:

kubectl get ingress envoy

and re-apply the kube.yaml with {TUNNEL_DOMAIN} substitution again.

Test Result on GKE asia-east1-a

3 backend pods + 9 client pods -> direct GCP Loadbalancer

10K requests/s. In this test case, the lower latency and higher throughput is expected, because the clients and the server pods are all in the GCP network.