Benchmark Operator

The intent of this Operator is to deploy common workloads to establish a performance baseline of Kubernetes cluster on your provider.

Installation (Default)

The easiest way to install the operator is through the operator-sdk methods provided in the Makefile.

git clone https://github.com/cloud-bulldozer/benchmark-operator
cd benchmark-operator
make deploy

If you wish to build a version of the operator from your local copy of the repo, you can run

git clone https://github.com/cloud-bulldozer/benchmark-operator
cd benchmark-operator
make image-build image-push deploy IMG=$YOUR_IMAGE

Note: building the image requires podman

Installation (Helm)

Installing the benchmark-operator via Helm can be done with the following commands. This requires your machine to have Helm installed. Install Helm

Note: If running on openshift you'll need to run this command before installing the chart. oc adm policy -n benchmark-operator add-scc-to-user privileged -z benchmark-operator

git clone https://github.com/cloud-bulldozer/benchmark-operator
cd benchmark-operator/charts/benchmark-operator
kubectl create namespace benchmark-operator 
oc adm policy -n benchmark-operator add-scc-to-user privileged -z benchmark-operator # Openshift Only 
helm install benchmark-operator . -n benchmark-operator --create-namespace

To delete this release, you can do so with the following command:

helm uninstall benchmark-operator -n benchmark-operator

Workloads status

Workload	Use	ElasticSearch indexing	Reconciliation usage	VM support (kubevirt)	Kata Containers	CI Tested
UPerf	Network Performance	Yes	Used, default : 3second	Working	Working	Yes
Iperf3	Network Performance	No	Used, default : 3second	Not Supported	Preview	Yes
fio	Storage IO	Yes	Used, default : 3second	Working	Working	Yes
Sysbench	System Performance	No	Used, default : 3second	Not Supported	Preview	Yes
YCSB	Database Performance	Yes	Used, default : 3second	Not Supported	Preview	Yes
Byowl	User defined workload	Yes	Used, default : 3second	Not Supported	Preview	Yes
Pgbench	Postgres Performance	Yes	Used, default : 3second	Not Supported	Preview	Yes
Smallfile	Storage IO Performance	Yes	Used, default : 3second	Not Supported	Preview	Yes
fs-drift	Storage IO Longevity	Yes	Not used	Not Supported	Preview	Yes
hammerdb	Database Performance	Yes	Used, default : 3second	Working	Preview	Yes
Service Mesh	Microservices	No	Used, default : 3second	Not Supported	Preview	No
Vegeta	HTTP Performance	Yes	Used, default : 3second	Not Supported	Preview	Yes
Scale Openshift	Scale Openshift Cluster	Yes	Used, default : 3second	Not Supported	Preview	Yes
stressng	Stress system resources	Yes	Used, default: 3second	Working	Preview	Yes
kube-burner	k8s Performance	Yes	Used, default : 3second	Not Supported	Preview	Yes
cyclictest	Real-Time Performance	Yes	Used, default : 3second	Not Supported	Preview	No
oslat	Real-Time Latency	Yes	Used, default : 3second	Not Supported	Preview	No
testpmd	TestPMD DPDK App	No	Used	Not Supported	Preview	No
Flent	Network Performance	Yes	Used, default : 3second	Not Supported	Not Supported	Yes
Log-Generator	Log Throughput to Backend	Yes	Used, default : 3second	Not Supported	Yes	Yes
Image-Pull	Time to Pull Image from Container Repo	Yes	Used, default : 3second	Not Supported	Yes	Yes

Reconciliation

Previously the Benchmark Operator didn't properly take advantage of the reconciliation period. Going forward we will make every attempt to utilize the reconciliation period.

Why did we decide to switch to this? Our operator would implement long running tasks, due to the nature of benchmarks. However, long running tasks blocks the Operator, causing us to delete the Operator and re-create the operator to un-block it. The benchmarks mentioned above that state Used for Reconciliation, no longer have this issue.

Optional workload images

Optional locations for workload images can now be added easily without the need to rebuild the operator. To do so in the workload args section of the CR add image: [location]

NOTE: This is not a required arguement. If omitted it will default to the currently verified workload image. Additionally, this is NOT enabled for YCSB

For Example:

apiVersion: ripsaw.cloudbulldozer.io/v1alpha1
kind: Benchmark
metadata:
  name: example-benchmark
  namespace: benchmark-operator
spec:
  elasticsearch:
    url: "http://my-es.foo.bar:80"
  metadata_collection: true
  cleanup: false
  workload:
    name: "foo"
    args:
      image: my.location/foo:latest

Optional debug out for benchmark-wrapper workloads

Workloads that are triggered through benchmark-wrapper can optionally pass the debug flag through the workload CR.

NOTE: This is not a required arguement. If omitted it will default to the default logging level of the benchmark-wrapper

For Example:

apiVersion: ripsaw.cloudbulldozer.io/v1alpha1
kind: Benchmark
metadata:
  name: example-benchmark
  namespace: benchmark-operator
spec:
  elasticsearch:
    url: "http://my-es.foo.bar:80"
  metadata_collection: true
  cleanup: false
  workload:
    name: snafu_workload
    args:
      debug: true

User Provided UUID

All benchmarks in the benchmark-operator utilize a UUID for tracking and indexing purposes. This UUID is, by default, generated when the workload is first started. However, if desired, a user provided UUID can be added to the workload cr.

NOTE: The provided UUID must be in format XXXXX-XXXXX-XXXXX

For Example:

apiVersion: ripsaw.cloudbulldozer.io/v1alpha1
kind: Benchmark
metadata:
  name: example-benchmark
  namespace: benchmark-operator
spec:
  uuid: 6060004a-7515-424e-93bb-c49844600dde
  elasticsearch:
    url: "http://my-es.foo.bar:80"
  metadata_collection: true
  cleanup: false
  workload:
    name: "foo"
    args:
      image: my.location/foo:latest

Contributing

Contributing

Metadata Collection

Metadata Collection

Indexing to Elasticsearch

Indexing to Elasticsearch

Capturing Prometheus Data

Capturing Prometheus Data

Cache dropping

Cache dropping

Community

Key Members(slack_usernames): ravi, mohit, dry923, rsevilla or rook

• #sig-scalability on Kubernetes Slack
• #forum-perfscale on CoreOS Slack