gRPC Load Balancing Demo

A simple demo of gRPC load balancing.

Environment requirements

The following software is required to run the demo:

• Go 1.14 or above.
• Kubernetes
• Skaffold 1.6.0 or above.
• Linkerd 2.0 or above.

Demo in native mode

Build

Build the native binaries:

./build.sh

The following binaries will be generated in the out directory:

• server: sends its IP address back to HTTP and gRPC clients.

• client-http: connects to server via HTTP.

• client-grpc: connects to server via gRPC.

Run

Arrange your terminal as follows for better experience:

Pane A
Start the server:

out/server

The server provides the same addr service via 2 TCP ports:

• Port 80 for http
• Port 30051 for gRPC

Pane B
Connect to server’s addr service via http endpoint:

out/client-http  http://127.0.0.1:80/addr

You can see the IP address of the connected server instance.

Pane C
Connect to server’s addr service via gRPC endpoint:

out/client-grpc  localhost:30051

You can see the IP address of the connected server instance.

Demo in Kubernetes

Preparation

Create grpc-lb namespace for this demo:

kubectl create ns grpc-lb

Build images:

skaffold build

Run without gRPC load balancing

Run server (only one pod) in grpc-lb namespace:

skaffold dev  -n grpc-lb

The server provides the same addr service via 2 TCP ports:

• Port 30080 for http
• Port 30051 for gRPC

Run native client-http:

out/client-http  http://127.0.0.1:30080/addr

Run native client-grpc:

out/client-grpc  localhost:30051

You can see the IP address of the single server pod is 10.1.0.8:

Now, scale the server to 5 pods:

kubectl scale -n grpc-lb --replicas=5 deployment/addr-server

You can see that only one of the gRPC server pods is being connected to and serving the same client-grpc instance:

Demo in Kubernetes + service mesh

Enable Linkerd

Install Linkerd 2 into active Kubernetes cluster:

# Install
linkerd install | kubectl apply -f -
# Check
linkerd check

Delete the old namespace grpc-lb:

kubectl delete ns grpc-lb

Create a new namespace grpc-lb with Linkerd injected:

kubectl apply -f ns.yml

Check data plane again, if necessary:

linkerd -n grpc-lb check --proxy

Run with gRPC load balancing

Uncomment the following line in skaffold.yaml:

      - client-grpc.yml

Run server and k8s version of client-grpc in grpc-lb namespace:

skaffold dev  -n grpc-lb

Run native client-grpc, as a comparison:

out/client-grpc  localhost:30051

Make sure you have run these commands as follows:

Wait for the whole clients and server becoming stable.

Now, scale the server to 5 pods:

kubectl scale -n grpc-lb --replicas=5 deployment/addr-server

You can see that all gRPC server pods (10.1.0.71 — 10.1.0.75) take turns in serving the same client instance within the same grpc-lb namespace.

View Linkerd dashboard:

linkerd dashboard &

You can see the topology of native client-grpc, meshed client-grpc and meshed server:

You can also see that all gRPC server pods take turns in serving the same client instance within the same grpc-lb namespace.