kafka_performance

README

Performance test image

This image is designed to benchmark Knative Eventing channel/brokers.

The image contains three different functionality, configurable with --roles flag:

• sender: Act as sender
• receiver: Act as receiver
• aggregator: Act as aggregator of timestamps

The image does both the sender and receiver role, allowing the clock to be synchronized to correctly calculate latencies (only valid with a single sender-receiver).

Latencies are calculated and published to the Mako sidecar container by a separate aggregator. In particular:

• Publish latency represents how long the sender took to publish the event to its sink
• Delivery latency represents how much long the event took to go from the "intent" of publish (before publishing request happens) to the delivery of the event in the receiver

Throughputs are calculated doing the finite difference of timestamp - count vector. In particular:

• Send throughput are calculated from the timestamps stored before the publish request
• Delivery throughput are calculated from the timestamps stored when the message are received
• Failures throughput are calculated from the timestamps / failures vector

The image is designed to allocate as much memory as possible before the benchmark starts. We suggest to disable Go GC to avoid useless GC pauses.

Usage

Example of how to use this image with the Mako stub sidecar:

apiVersion: v1
kind: Pod
metadata:
  name: latency-test
  namespace: perf-eventing
  labels:
    role: latency-test-consumer
spec:
  serviceAccountName: default
  restartPolicy: Never
  containers:
    - name: latency-test
      image: knative.dev/eventing/test/test_images/performance
      resources:
        requests:
          cpu: 1000m
          memory: 2Gi
      ports:
        - name: cloudevents
          containerPort: 8080
      args:
        - "--roles=sender,receiver"
        - "--sink=http://in-memory-test-broker-broker.perf-eventing.svc.cluster.local"
        - "--aggregator=localhost:10000"
        - "--pace=100:10,200:20,400:60"
        - "--warmup=10"
      env:
        - name: POD_NAMESPACE
          valueFrom:
            fieldRef:
              fieldPath: metadata.namespace
      volumeMounts:
        - name: config-mako
          mountPath: /etc/config-mako
      terminationMessagePolicy: FallbackToLogsOnError
    - name: aggregator
      image: knative.dev/eventing/test/test_images/performanceaggregator
      ports:
        - name: grpc
          containerPort: 10000
      args:
        - "--roles=aggregator"
        - "--expect-records=2"
      terminationMessagePolicy: FallbackToLogsOnError
    - name: mako-stub
      image: knative.dev/pkg/test/mako/stub-sidecar
      terminationMessagePolicy: FallbackToLogsOnError
  volumes:
    - name: config-mako
      configMap:
        name: config-mako

Pace configuration

pace is a comma separated array of pace configurations in format rps[:duration=10s].

For example the configuration 100,200:20,400:60 means:

1. 100 rps for 10 seconds
2. 200 rps for 20 seconds
3. 400 rps for 60 seconds

Warmup phase

You can configure a warmup phase to warm the hot path of channel implementations. This is especially required while working with JVM or similar environments. During the warmup phase, no latencies are calculated.

To configure the duration of warmup phase, use flag warmup specifying the number of seconds.

If you don't want a warmup phase, use --warmup=0.

Workers

You can specify the number of initial vegeta workers that perform requests with flag workers.

Aggregator configuration

--expect-records must be equal to number sender + number receivers. If a same instance does both the sender and receiver, it counts twice