Reproduce the two-tier HTTP benchmark Environment Dedicated, rootful Lima VM on Apple silicon. Ubuntu 26.04.1 LTS (aarch64), kernel 7.0.0-43-generic, 4 vCPU, 4 GiB configured memory. Local disk, a single k3s server/agent, default flannel VXLAN, CoreDNS. The HTTP client runs on the node. TLS, WAN transit, a database and an AWS load balancer are outside the traffic path. Caches stay warm within a block. The host is shared. Four observations per runtime are descriptive. Versions and identities k3s v1.37.2+k3s1, upstream k3s-arm64 asset, SHA-256 a779ccf1e5e2e33a62f5a29c2f1c0bbf99431e100579d4c10f0aa0742a6aa84b https://github.com/k3s-io/k3s/releases/tag/v1.37.2%2Bk3s1 Docker Engine 29.1.3, Ubuntu build 29.1.3-0ubuntu4.1. Podbag release-profile arm64 build of c0ff5ebdbd8762760652fbe0a869655a105557a9 SHA-256 b2c5242d66ec1b1504fcabd7dd6dc02b8f872427e75fece25f366ab4d8a286eb https://github.com/SteelCrab/podbag/tree/c0ff5ebdbd8762760652fbe0a869655a105557a9 NGINX 1.28.3-alpine, linux/arm64 manifest: docker.io/library/nginx@sha256:1be286f7dd7d6e6be04f46e8df36d0084b05aa3d7fd3c6ebda3e1eff4e5b665d Files workload.json contains two ConfigMaps, two ClusterIP Services, two Deployments. start-node.sh starts one runtime and a fresh k3s cluster per block. measure.py creates the workload at zero replicas, warms it once, then measures two runs. It owns the namespace shop-bench in this disposable VM. results.jsonl includes all four warm-ups and all eight timed observations. Node preparation Use a disposable Linux arm64 VM with systemd, cgroup v2, Docker, crictl, Python 3, curl and erofs-utils. Put k3s and podbag in /usr/local/bin. Use the cgroupfs driver for Docker and kubelet. Configure the Docker daemon with {"exec-opts":["native.cgroupdriver=cgroupfs"]}. Disable automatic startup of docker.service, docker.socket and containerd.service. start-node.sh starts Docker only in Docker blocks. Reserve the VM for the benchmark. Save workload.json, start-node.sh and measure.py in one directory. Reboot the same VM before each block. The cluster state is new; the image stores can be reused. Run the following inside the VM. Use sudo -i and enter the file directory first. Use (runtime,block) in this order: (docker,1) (podbag,2) (podbag,3) (docker,4). Set rt and block for the current block: rt=docker block=1 bash start-node.sh "$rt" "$block" image=docker.io/library/nginx@sha256:1be286f7dd7d6e6be04f46e8df36d0084b05aa3d7fd3c6ebda3e1eff4e5b665d sock=/run/pb30/cri.sock if [ "$rt" = docker ]; then docker pull "$image" sock=$(find /run/k3s -type s -name '*cri*dock*.sock' | head -1) fi cri() { crictl --config /dev/null -r "unix://$sock" -i "unix://$sock" "$@"; } cri inspecti "$image" >/dev/null || cri pull "$image" cri inspecti "$image" # verify the full reference in status.repoDigests k3s kubectl -n kube-system wait --for=condition=complete job/helm-install-gateway-api-crd --timeout=240s k3s kubectl get node bench -o jsonpath='{.status.nodeInfo.containerRuntimeVersion}' python3 measure.py "$rt" "$block" | tee "block-$block.jsonl" k3s kubectl -n kube-system scale deploy/coredns --replicas=0 k3s kubectl -n kube-system wait --for=delete pod -l k8s-app=kube-dns --timeout=40s sync Metrics ready_s: wall time from just before patching API then web replica counts until both Deployments report 24+6 Ready, polled at 100 ms intervals. Each trial starts after all pods from the previous trial disappear. System pods are extra. The client then checks the two EndpointSlices and waits two seconds. burst: 12 clients x 100 requests, no pacing, fresh TCP connection each time. steady: after two more seconds, 12 clients x 600 requests spaced 50 ms apart per client (target 240 requests/s for 30 s). The caller uses the web ClusterIP. success: HTTP 200 and a response beginning with the API pod name prefix. backends: count of distinct API pod names in valid responses. p95_ms: nearest-rank 95th percentile within a run. successful_rps: number of validated responses divided by the wall time of the load phase. Article table: unweighted medians of the four timed results for each metric. The phases are separate and each warm-up (run=0) is excluded. The validated response row adds both phases across all four timed trials per runtime. Image reference check from the preliminary run The multi-architecture index digest for the same NGINX image was sha256:a8b39bd9cf0f83869a2162827a0caf6137ddf759d50a171451b335cecc87d236. With IfNotPresent and that reference, the Podbag CRI listed the arm64 manifest among repoDigests; inspecti by index reference returned "no such image" and kubelet reported "pull QPS exceeded" for retrying pods. The main experiment pins the arm64 manifest and uses Never on every application pod. On the Podbag pilot node the following two lookups distinguished the two references to the arm64 content. Run these after pulling the index: cri inspecti docker.io/library/nginx@sha256:a8b39bd9cf0f83869a2162827a0caf6137ddf759d50a171451b335cecc87d236 cri inspecti docker.io/library/nginx@sha256:1be286f7dd7d6e6be04f46e8df36d0084b05aa3d7fd3c6ebda3e1eff4e5b665d In a separate rehearsal, the CoreDNS pod from a previously booted cluster entered CrashLoopBackOff; web proxies could not resolve the API Service. That rehearsal is excluded. The recorded blocks each start a fresh cluster and wait for CoreDNS before deploying the application.