node-performance-testing

Measuring Node Performance

This document outlines the issues and pitfalls of measuring Node performance, as well as the tools available.

Cluster Set-up

There are lots of factors which can affect node performance numbers, so care must be taken in setting up the cluster to make the intended measurements. In addition to taking the following steps into consideration, it is important to document precisely which setup was used. For example, performance can vary wildly from commit-to-commit, so it is very important to document which commit or version of Kubernetes was used, which Docker version was used, etc.

Addon pods

Be aware of which addon pods are running on which nodes. By default Kubernetes runs 8 addon pods, plus another 2 per node (fluentd-elasticsearch and kube-proxy) in the kube-system namespace. The addon pods can be disabled for more consistent results, but doing so can also have performance implications.

For example, Heapster polls each node regularly to collect stats data. Disabling Heapster will hide the performance cost of serving those stats in the Kubelet.

Disabling Add-ons

Disabling addons is simple. Just ssh into the Kubernetes master and move the addon from /etc/kubernetes/addons/ to a backup location. More details here.

Which / how many pods?

Performance will vary a lot between a node with 0 pods and a node with 100 pods. In many cases you’ll want to make measurements with several different amounts of pods. On a single node cluster scaling a replication controller makes this easy, just make sure the system reaches a steady-state before starting the measurement. E.g. kubectl scale replicationcontroller pause --replicas=100

In most cases pause pods will yield the most consistent measurements since the system will not be affected by pod load. However, in some special cases Kubernetes has been tuned to optimize pods that are not doing anything, such as the cAdvisor housekeeping (stats gathering). In these cases, performing a very light task (such as a simple network ping) can make a difference.

Finally, you should also consider which features yours pods should be using. For example, if you want to measure performance with probing, you should obviously use pods with liveness or readiness probes configured. Likewise for volumes, number of containers, etc.

Other Tips

Number of nodes - On the one hand, it can be easier to manage logs, pods, environment etc. with a single node to worry about. On the other hand, having multiple nodes will let you gather more data in parallel for more robust sampling.

E2E Performance Test

There is an end-to-end test for collecting overall resource usage of node components: kubelet_perf.go. To run the test, simply make sure you have an e2e cluster running (go run hack/e2e.go -- -up) and set up correctly.

Run the test with go run hack/e2e.go -- -v -test --test_args="--ginkgo.focus=resource\susage\stracking". You may also wish to customise the number of pods or other parameters of the test (remember to rerun make WHAT=test/e2e/e2e.test after you do).

Profiling

Kubelet installs the go pprof handlers, which can be queried for CPU profiles:

$ kubectl proxy &
Starting to serve on 127.0.0.1:8001
$ curl -G "http://localhost:8001/api/v1/proxy/nodes/${NODE}:10250/debug/pprof/profile?seconds=${DURATION_SECONDS}" > $OUTPUT
$ KUBELET_BIN=_output/dockerized/bin/linux/amd64/kubelet
$ go tool pprof -web $KUBELET_BIN $OUTPUT

pprof can also provide heap usage, from the /debug/pprof/heap endpoint (e.g. http://localhost:8001/api/v1/proxy/nodes/${NODE}:10250/debug/pprof/heap).

More information on go profiling can be found here.

Benchmarks

Before jumping through all the hoops to measure a live Kubernetes node in a real cluster, it is worth considering whether the data you need can be gathered through a Benchmark test. Go provides a really simple benchmarking mechanism, just add a unit test of the form:

// In foo_test.go
func BenchmarkFoo(b *testing.B) {
  b.StopTimer()
  setupFoo() // Perform any global setup
  b.StartTimer()
  for i := 0; i < b.N; i++ {
    foo() // Functionality to measure
  }
}

Then:

$ go test -bench=. -benchtime=${SECONDS}s foo_test.go

More details on benchmarking here.

TODO

  • (taotao) Measuring docker performance
  • Expand cluster set-up section
  • (vishh) Measuring disk usage
  • (yujuhong) Measuring memory usage
  • Add section on monitoring kubelet metrics (e.g. with prometheus)