feat: Add rolling avg toggle (#2)

* Add rolling avg feature

Signed-off-by: Erez Zarum <erezz@amazon.com>

* Fix cargo fmt issues, add extra metrics selection

Signed-off-by: Erez Zarum <erezz@amazon.com>

* Add kubernetes example

Signed-off-by: Erez Zarum <erezz@amazon.com>

* Implement fixes

Signed-off-by: Erez Zarum <erezz@amazon.com>

---------

Signed-off-by: Erez Zarum <erezz@amazon.com>

Author: erezzarum

README.md

@@ -47,6 +47,7 @@ communication benchmarks:
 - [NVSHMEM](examples/nvshmem/) — one-sided GPU communication
 - [PPLX Kernels](examples/pplx/) — MoE all-to-all dispatch/combine
 - [RDMA Statistics](examples/rdma/) — shell-based RDMA stats
+- [Kubernetes](examples/kubernetes/) — DaemonSet deployment for Kubernetes
 
 ## How It Works
 

examples/kubernetes/Dockerfile

@@ -0,0 +1,10 @@
+FROM rust:1-alpine AS builder
+RUN apk add --no-cache musl-dev
+WORKDIR /build
+COPY Cargo.toml ./
+COPY src/ src/
+RUN cargo build --release
+
+FROM alpine:3
+COPY --from=builder /build/target/release/rdmatop /usr/local/bin/rdmatop
+ENTRYPOINT ["/usr/local/bin/rdmatop"]

examples/kubernetes/README.md

@@ -0,0 +1,57 @@
+# Kubernetes
+
+Deploy `rdmatop` as a DaemonSet to monitor RDMA traffic across all
+GPU nodes in a Kubernetes cluster. The DaemonSet runs a container with
+`NET_ADMIN` capability and host PID/network namespaces on each matching
+node, giving you per-node RDMA visibility without SSH access.
+
+## Build
+
+Build the container image from the project root:
+
+```bash
+cd rdmatop
+docker build -f examples/kubernetes/Dockerfile -t rdmatop:latest .
+```
+
+Push to your registry if needed:
+
+```bash
+docker tag rdmatop:latest <registry>/rdmatop:latest
+docker push <registry>/rdmatop:latest
+```
+
+## Deploy
+
+Before deploying, update the `image` field in `daemonset.yaml` to point to your registry (e.g. `<registry>/rdmatop:latest`) so Kubernetes can pull the image:
+
+```bash
+kubectl apply -f examples/kubernetes/daemonset.yaml
+```
+
+The DaemonSet targets GPU instance types (p5, p5e, p5en, p6e, p6-b200,
+p6-b300) via node affinity. It uses `hostNetwork: true` so the container
+sees the host's RDMA devices, `hostPID: true` so rdmatop can read
+`/proc/<pid>` to enrich RDMA QP owners, and requests `NET_ADMIN`
+capability for netlink access.
+
+## Usage
+
+Exec into a running pod to launch `rdmatop`:
+
+```bash
+# List pods
+kubectl get pods -l app=rdmatop
+
+# Attach to a specific node's pod
+kubectl exec -it <pod-name> -- rdmatop
+```
+
+## Configuration
+
+| Field | Description |
+|-------|-------------|
+| `hostNetwork: true` | Required — exposes host RDMA devices to the container |
+| `hostPID: true` | Required — allows rdmatop to read `/proc/<pid>` for RDMA process info |
+| `NET_ADMIN` | Required — allows netlink queries for RDMA stats |
+| `nodeAffinity` | Targets GPU instance types; edit `values` to match your cluster |

examples/kubernetes/daemonset.yaml

@@ -0,0 +1,42 @@
+apiVersion: apps/v1
+kind: DaemonSet
+metadata:
+  name: rdmatop
+  labels:
+    app: rdmatop
+spec:
+  selector:
+    matchLabels:
+      app: rdmatop
+  template:
+    metadata:
+      labels:
+        app: rdmatop
+    spec:
+      affinity:
+        nodeAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+            nodeSelectorTerms:
+            - matchExpressions:
+              - key: node.kubernetes.io/instance-type
+                operator: In
+                values:
+                - p5.48xlarge
+                - p5e.48xlarge
+                - p5en.48xlarge
+                - p6e-gb200.36xlarge
+                - p6-b200.48xlarge
+                - p6-b300.48xlarge
+      hostNetwork: true
+      hostPID: true
+      containers:
+        - name: rdmatop
+          image: rdmatop:latest
+          imagePullPolicy: IfNotPresent
+          command: ["sleep", "infinity"]
+          securityContext:
+            capabilities:
+              add:
+                - NET_ADMIN
+      tolerations:
+        - operator: Exists