In this article, you will learn how to use virtual threads, build a native image with GraalVM and run such the Java app on Kubernetes. Currently, the native compilation (GraalVM) and virtual threads (Project Loom) are probably the hottest topics in the Java world. They improve the general performance of your app including memory usage and startup time. Since startup time and memory usage were always a problem for Java, expectations for native images or virtual threads were really big.

Of course, we usually consider such performance issues within the context of microservices or serverless apps. They should not consume many OS resources and should be easily auto-scalable. We can easily control resource usage on Kubernetes. If you are interested in Java virtual threads you can read my previous article about using them to create an HTTP server available here. For more details about Knative as serverless on Kubernetes, you can refer to the following article.

Introduction

Let’s start with the plan for our exercise today. In the first step, we will create a simple Java web app that uses virtual threads for processing incoming HTTP requests. Before we run the sample app we will install Knative on Kubernetes to quickly test autoscaling based on HTTP traffic. We will also install Prometheus on Kubernetes. This monitoring stack allows us to compare the performance of the app without/with GraalVM and virtual threads on Kubernetes. Then, we can proceed with the deployment. In order to easily build and run our native app on Kubernetes we will use Cloud Native Buildpacks. Finally, we will perform some load tests and compare metrics.

Source Code

If you would like to try it by yourself, you may always take a look at my source code. In order to do that you need to clone my GitHub repository. After that, you should follow my instructions.

Create Java App with Virtual Threads

In the first step, we will create a simple Java app that acts as an HTTP server and handles incoming requests. In order to do that, we can use the HttpServer object from the core Java API. Once we create the server we can override a default thread executor with the setExecutor method. In the end, we will try to compare the app using standard threads with the same app using virtual threads. Therefore, we allow overriding the type of executor using an environment variable. The name of that is THREAD_TYPE. If you want to enable virtual threads you need to set the value virtual for that env. Here’s the main method of our app.

public class MainApp {

   public static void main(String[] args) throws IOException {
      HttpServer httpServer = HttpServer
         .create(new InetSocketAddress(8080), 0);

      httpServer.createContext("/example", 
         new SimpleCPUConsumeHandler());

      if (System.getenv("THREAD_TYPE").equals("virtual")) {
         httpServer.setExecutor(
            Executors.newVirtualThreadPerTaskExecutor());
      } else {
         httpServer.setExecutor(Executors.newFixedThreadPool(200));
      }
      httpServer.start();
   }

}

In order to process incoming requests, the HTTP server uses the handler that implements the HttpHandler interface. In our case, the handler is implemented inside the SimpleCPUConsumeHandler class as shown below. It consumes a lot of CPU since it creates an instance of BigInteger with the constructor that performs a lot of computations under the hood. It will also consume some time, so we have the simulation of processing time in the same step. As a response, we just return the next number in the sequence with the Hello_ prefix.

public class SimpleCPUConsumeHandler implements HttpHandler {

   Logger LOG = Logger.getLogger("handler");
   AtomicLong i = new AtomicLong();
   final Integer cpus = Runtime.getRuntime().availableProcessors();

   @Override
   public void handle(HttpExchange exchange) throws IOException {
      new BigInteger(1000, 3, new Random());
      String response = "Hello_" + i.incrementAndGet();
      LOG.log(Level.INFO, "(CPU->{0}) {1}", 
         new Object[] {cpus, response});
      exchange.sendResponseHeaders(200, response.length());
      OutputStream os = exchange.getResponseBody();
      os.write(response.getBytes());
      os.close();
   }
}

In order to use virtual threads in Java 19 we need to enable preview mode during compilation. With Maven we need to enable preview features using maven-compiler-plugin as shown below.

<plugin>
  <groupId>org.apache.maven.plugins</groupId>
  <artifactId>maven-compiler-plugin</artifactId>
  <version>3.10.1</version>
  <configuration>
    <release>19</release>
    <compilerArgs>
      --enable-preview
    </compilerArgs>
  </configuration>
</plugin>

Install Knative on Kubernetes

This and the next step are not required to run the native application on Kubernetes. We will use Knative to easily autoscale the app in reaction to the volume of incoming traffic. In the next section, I’ll describe how to run a monitoring stack on Kubernetes.

The simplest way to install Knative on Kubernetes is with the kubectl command. We just need the Knative Serving component without any additional features. The Knative CLI (kn) is not required. We will deploy the application from the YAML manifest using Skaffold.

First, let’s install the required custom resources with the following command:

$ kubectl apply -f https://github.com/knative/serving/releases/download/knative-v1.8.3/serving-crds.yaml

Then, we can Install the core components of Knative Serving by running the command:

$ kubectl apply -f https://github.com/knative/serving/releases/download/knative-v1.8.3/serving-core.yaml

In order to access Knative services outside of the Kubernetes cluster we also need to install a networking layer. By default, Knative uses Kourier as an ingress. We can install the Kourier controller by running the following command.

$ kubectl apply -f https://github.com/knative/net-kourier/releases/download/knative-v1.8.1/kourier.yaml

Finally, let’s configure Knative Serving to use Kourier with the following command:

kubectl patch configmap/config-network \
  --namespace knative-serving \
  --type merge \
  --patch '{"data":{"ingress-class":"kourier.ingress.networking.knative.dev"}}'

If you don’t have an external domain configured or you are running Knative on the local cluster you need to configure DNS. Otherwise, you would have to run curl commands with a host header. Knative provides a Kubernetes Job that sets sslip.io as the default DNS suffix.

$ kubectl apply -f https://github.com/knative/serving/releases/download/knative-v1.8.3/serving-default-domain.yaml

The generated URL contains the name of the service, the namespace, and the address of your Kubernetes cluster. Since I’m running my service on the local Kubernetes cluster in the demo-sless namespace my service is available under the following address:

But before we deploy the sample app on Knative, let’s do some other things.

Install Prometheus Stack on Kubernetes

As I mentioned before, we can also install a monitoring stack on Kubernetes.

The simplest way to install it is with the kube-prometheus-stack Helm chart. The package contains Prometheus and Grafana. It also includes all required rules and dashboards to visualize the basic metrics of your Kubernetes cluster. Firstly, let’s add the Helm repository containing our chart:

$ helm repo add prometheus-community https://prometheus-community.github.io/helm-charts

Then we can install the kube-prometheus-stack Helm chart in the prometheus namespace with the following command:

$ helm install prometheus-stack prometheus-community/kube-prometheus-stack  \
    -n prometheus \
    --create-namespace

If everything goes fine, you should see a similar list of Kubernetes services:

$ kubectl get svc -n prometheus
NAME                                        TYPE        CLUSTER-IP       EXTERNAL-IP   PORT(S)                      AGE
alertmanager-operated                       ClusterIP   None             <none>        9093/TCP,9094/TCP,9094/UDP   11s
prometheus-operated                         ClusterIP   None             <none>        9090/TCP                     10s
prometheus-stack-grafana                    ClusterIP   10.96.218.142    <none>        80/TCP                       23s
prometheus-stack-kube-prom-alertmanager     ClusterIP   10.105.10.183    <none>        9093/TCP                     23s
prometheus-stack-kube-prom-operator         ClusterIP   10.98.190.230    <none>        443/TCP                      23s
prometheus-stack-kube-prom-prometheus       ClusterIP   10.111.158.146   <none>        9090/TCP                     23s
prometheus-stack-kube-state-metrics         ClusterIP   10.100.111.196   <none>        8080/TCP                     23s
prometheus-stack-prometheus-node-exporter   ClusterIP   10.102.39.238    <none>        9100/TCP                     23s

We will analyze Grafana dashboards with memory and CPU statistics. We can enable port-forward to access it locally on the defined port, for example 9080:

$ kubectl port-forward svc/prometheus-stack-grafana 9080:80 -n prometheus

The default username for Grafana is admin and password prom-operator.

We will create two panels in the custom Grafana dashboard. First of them will show the memory usage per single pod in the demo-sless namespace.

sum(container_memory_working_set_bytes{namespace="demo-sless"} / (1024 * 1024)) by (pod)

The second of them will show the average CPU usage per single pod in the demo-sless namespace. You can import both of these directly to Grafana from the k8s/grafana-dasboards.json file from the GitHub repo.

rate(container_cpu_usage_seconds_total{namespace="demo-sless"}[3m])

Build and Deploy a native Java Application

We have already created the sample app and then configured the Kubernetes environment. Now, we may proceed to the deployment phase. Our goal here is to simplify the process of building a native image and running it on Kubernetes as much as possible. Therefore, we will use Cloud Native Buildpacks and Skaffold. With Buildpacks we don’t need to have anything installed on our laptop besides Docker. Skaffold can be easily integrated with Buildpacks to automate the whole process of building and running the app on Kubernetes. You just need to install the skaffold CLI on your machine.

For building a native image of a Java application we may use Paketo Buildpacks. It provides a dedicated buildpack for GraalVM called Paketo GraalVM Buildpack. We should include it in the configuration using the paketo-buildpacks/graalvm name. Since Skaffold supports Buildpacks, we should set all the properties inside the skaffold.yaml file. We need to override some default settings with environment variables. First of all, we have to set the version of Java to 19 and enable preview features (virtual threads). The Kubernetes deployment manifest is available under the k8s/deployment.yaml path.

apiVersion: skaffold/v2beta29
kind: Config
metadata:
  name: sample-java-concurrency
build:
  artifacts:
  - image: piomin/sample-java-concurrency
    buildpacks:
      builder: paketobuildpacks/builder:base
      buildpacks:
        - paketo-buildpacks/graalvm
        - paketo-buildpacks/java-native-image
      env:
        - BP_NATIVE_IMAGE=true
        - BP_JVM_VERSION=19
        - BP_NATIVE_IMAGE_BUILD_ARGUMENTS=--enable-preview
  local:
    push: true
deploy:
  kubectl:
    manifests:
    - k8s/deployment.yaml

Knative simplifies not only autoscaling, but also Kubernetes manifests. Here’s the manifest for our sample app available in the k8s/deployment.yaml file. We need to define a single object Service containing details of the application container. We will change the autoscaling target from the default 200 concurrent requests to 80. It means that if a single instance of the app will process more than 80 requests simultaneously Knative will create a new instance of the app (or a pod – to be more precise). In order to enable virtual threads for our app we also need to set the environment variable THREAD_TYPE to virtual.

apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: sample-java-concurrency
spec:
  template:
    metadata:
      annotations:
        autoscaling.knative.dev/target: "80"
    spec:
      containers:
        - name: sample-java-concurrency
          image: piomin/sample-java-concurrency
          ports:
            - containerPort: 8080
          env:
            - name: THREAD_TYPE
              value: virtual
            - name: JAVA_TOOL_OPTIONS
              value: --enable-preview

Assuming you already installed Skaffold, the only thing you need to do is to run the following command:

$ skaffold run -n demo-sless

Or you can just deploy a ready image from my registry on Docker Hub. However, in that case, you need to change the image tag in the deployment.yaml manifest to virtual-native.

Once you deploy the app you can verify a list of Knative Service. The name of our target service is sample-java-concurrency. The address of the service is returned in the URL field.

$ kn service list -n demo-sless

Load Testing

We will run three testing scenarios today. In the first of them, we will test a standard compilation and a standard thread pool of 100 size. In the second of them, we will test a standard compilation with virtual threads. The final test will check native compilation in conjunction with virtual threads. In all these scenarios, we will set the same autoscaling target – 80 concurrent requests. I’m using the k6 tool for load tests. Each test scenario consists of 4 same steps. Each step takes 2 minutes. In the first step, we are simulating 50 users.

$ k6 run -u 50 -d 120s k6-test.js

Then, we are simulating 100 users.

$ k6 run -u 100 -d 120s k6-test.js

Finally, we run the test for 200 users twice. So, in total, there are four tests with 50, 100, 200, and 200 users, which takes 8 minutes.

$ k6 run -u 200 -d 120s k6-test.js

Let’s verify the results. By the way, here is our test for the k6 tool in javascript.

import http from 'k6/http';
import { check } from 'k6';

export default function () {
  const res = http.get(`http://sample-java-concurrency.demo-sless.127.0.0.1.sslip.io/example`);
  check(res, {
    'is status 200': (res) => res.status === 200,
    'body size is > 0': (r) => r.body.length > 0,
  });
}

Test for Standard Compilation and Threads

The diagram visible below shows memory usage at each phase of the test scenario. After simulating 200 users Knative scales up the number of instances. Theoretically, it should do that during 100 users test. But Knative measures incoming traffic at the level of the sidecar container inside the pod. The memory usage for the first instance is around ~900MB (it includes also sidecar container usage).

Here’s a similar view as before but for the CPU usage. The highest consumption was before autoscaling occurs at the level of ~1.2 core. Then, depending on the number of instances ranges from ~0.4 core to ~0.7 core. As I mentioned before, we are using a time-consuming BigInteger constructor to simulate CPU usage under a heavy load.

Here are the test results for 50 users. The application was able to process ~105k requests in 2 minutes. The highest processing time value was ~3 seconds.

Here are the test results for 100 users. The application was able to process ~130k requests in 2 minutes with an average response time of ~90ms.

Finally, we have results for 200 users test. The application was able to process ~135k requests in 2 minutes with an average response time of ~175ms. The failure threshold was at the level of 0.02%.

Test for Standard Compilation and Virtual Threads

The same as in the previous section, here’s the diagram that shows memory usage at each phase of the test scenario. After simulating 100 users Knative scales up the number of instances. Theoretically, it should run the third instance of the app for 200 users. The memory usage for the first instance is around ~850MB (it includes also sidecar container usage).

Here’s a similar view as before but for the CPU usage. The highest consumption was before autoscaling occurs at ~1.1 core. Then, depending on the number of instances ranges from ~0.3 core to ~0.7 core.

Here are the test results for 50 users. The application was able to process ~105k requests in 2 minutes. The highest processing time value was ~2.2 seconds.

Here are the test results for 100 users. The application was able to process ~115k requests in 2 minutes with an average response time of ~100ms.

Finally, we have results for 200 users test. The application was able to process ~135k requests in 2 minutes with an average response time of ~180ms. The failure threshold was at the level of 0.02%.

Test for Native Compilation and Virtual Threads

The same as in the previous section, here’s the diagram that shows memory usage at each phase of the test scenario. After simulating 100 users Knative scales up the number of instances. Theoretically, it should run the third instance of the app for 200 users (the third pod visible on the diagram was in fact in the Terminating phase for some time). The memory usage for the first instance is around ~50MB.

Here’s a similar view as before but for the CPU usage. The highest consumption was before autoscaling occurs at ~1.3 core. Then, depending on the number of instances ranges from ~0.3 core to ~0.9 core.

Here are the test results for 50 users. The application was able to process ~75k requests in 2 minutes. The highest processing time value was ~2 seconds.

Here are the test results for 100 users. The application was able to process ~85k requests in 2 minutes with an average response time of ~140ms

Finally, we have results for 200 users test. The application was able to process ~100k requests in 2 minutes with an average response time of ~240ms. Plus – there were no failures at the second 200 users attempt.

Summary

In this article, I tried to compare the behavior of the Java app for GraalVM native compilation with virtual threads on Kubernetes with a standard approach. There are several conclusions after running all described tests: