Article

Troubleshooting application performance with Red Hat OpenShift metrics, Part 3: Collecting runtime metrics

July 22, 2021
Kubernetes Operators
Pavel Macík Avatar
Pavel Macik
Principal Quality Engineer
Featured image for Operators topics.

This is the third article in series showing how to use metrics from Red Hat OpenShift to reveal application performance problems. In Part 1, I explained the environment and requirements for our application, the Service Binding Operator. In Part 2, I showed you how to set up a test environment in the Developer Sandbox for Red Hat OpenShift and introduced the test scenarios. Now, we can start to focus on the metrics themselves.

Read the whole series:

Collecting runtime metrics with the OpenShift monitoring tool

To see what is going on with Service Binding Operator and the OpenShift cluster under heavy load, it is important to collect metrics for the duration of the test. OpenShift's monitoring tool, a combination of Prometheus and Grafana with predefined, out-of-the-box metrics, can collect the necessary data for the lifespan of the cluster. At first, it seems an ideal solution to simply use that feature, let the monitoring stack gather the data normally, and collect it after the test is done. The problem is that all of the monitoring tools—the Prometheus and Graphana instances—are deployed on the same cluster as the application they're monitoring, including the OpenShift cluster's own resources. So, if the cluster goes down (which we expect to happen while stress testing), the monitoring subsystem goes down along with it, and all the gathered data is lost. Keep in mind that we are using a temporary development cluster that is terminated after about 10 hours anyway.

To ensure results are preserved on a node that won't crash, I've created the following collector script. It uses the OpenShift Client tool (oc) to pull the runtime metrics every 30 seconds or so for the duration of the test and stores it in a set of simple CSV files, one for nodes and one for each monitored pod. I start the script in the background before the user provisioning starts in order to catch the "before" state, and leave the script running for some time after the load generation ends to see the long-term behavior of the watched resources:

PERIOD="${1:-30}"
RESULTS=${2:-metrics-$(date "+%F_%T")}

mkdir -p "$RESULTS"

strip_unit(){
    echo -n $1 | sed -e 's,\([0-9]\+\)m,\1,g' | sed -e 's,\([0-9]\+\)Mi,\1,g' | sed -e 's,\([0-9]\+\)%,\1,g'
}

# Nodes
oc get nodes > $RESULTS/nodes.yaml
oc describe nodes > $RESULTS/nodes.info

node_info_file(){
    readlink -m "$RESULTS/node-info.$1.csv"
}

node_line(){
    node=$1
    node_json="$(oc get node $node -o json)"
    echo -n "$(echo "$node_json" | jq -rc '.status.conditions[] | select(.type=="MemoryPressure").status');"
    echo -n "$(echo "$node_json" | jq -rc '.status.conditions[] | select(.type=="DiskPressure").status');"
    echo -n "$(echo "$node_json" | jq -rc '.status.conditions[] | select(.type=="PIDPressure").status');"
    echo -n "$(echo "$node_json" | jq -rc '.status.conditions[] | select(.type=="Ready").status');"
    node_info=($(oc adm top node $node --no-headers))
    echo -n "$(strip_unit ${node_info[1]});"
    echo -n "$(strip_unit ${node_info[2]});"
    echo -n "$(strip_unit ${node_info[3]});"
    echo "$(strip_unit ${node_info[4]})"
}

NODES=($(oc get nodes -o json | jq -rc '.items[].metadata.name' | sort))
for node in "${NODES[@]}"; do
    echo "Time;MemoryPressure;DiskPressure;PIDPressure;Ready;CPU_millicores;CPU_percent;Memory_MiB;Memory_percent" > $(node_info_file $node)
done

# Operator pods
pod_info_file(){
    readlink -m "$RESULTS/pod-info.$1.csv"
}

pod_line(){
    pod=$1
    ns=$2
    pod_info=($(oc adm top pod $pod -n $ns --no-headers))
    echo -n "$(strip_unit ${pod_info[1]});"
    echo "$(strip_unit ${pod_info[2]})"
}

for namespace in openshift-operators openshift-monitoring openshift-apiserver openshift-kube-apiserver openshift-sdn openshift-operator-lifecycle-manager service-binding-operator; do
    PODS=($(oc get pods -n $namespace -o json | jq -rc '.items[].metadata.name' | grep -E 'operator|prometheus|apiserver|sdn|ovs|olm|packageserver' | sort))
    for pod in "${PODS[@]}"; do
        echo "Time;CPU_millicores;Memory_MiB" > $(pod_info_file $pod)
    done
done

echo "Collecting metrics"
# Periodical collection
while true; do
    echo -n "."
    for namespace in openshift-operators openshift-monitoring openshift-apiserver openshift-kube-apiserver openshift-sdn openshift-operator-lifecycle-manager service-binding-operator; do
        PODS=($(oc get pods -n $namespace -o json | jq -rc '.items[].metadata.name' | grep -E 'operator|prometheus|apiserver|sdn|ovs|olm|packageserver' | sort))
        for pod in "${PODS[@]}"; do
            pod_file=$(pod_info_file $pod)
            echo -n "$(date -u '+%F %T.%N');" >> $pod_file
            pod_line $pod $namespace >> $pod_file
        done
    done
    for node in ${NODES[@]}; do
        node_file=$(node_info_file $node)
        echo -n "$(date -u '+%F %T.%N');" >> $node_file
        node_line $node >> $node_file
    done
    sleep ${PERIOD}s
done

If the cluster survives the entire duration of the stress test, we can use Grafana to download the collected data from Prometheus in a form of similar CSV files.

Compiling a performance report

Since these tests were arranged pretty quickly and the metrics were collected in the raw form of CSV files, I had to manually convert the data into charts to put them into perspective. I used Google Sheets for that purpose.

Next steps

Finally, we have all the infrastructure we need for performance testing. In the next article, we will take our first look at the actual metrics and how they are gathered.

Read next: Part 4: Gathering performance metrics.

Last updated: September 19, 2023