The Performance Operator
optimizes OpenShift clusters for applications sensitive to cpu and network latency.
The operator was created using the operator-sdk:
$ operator-sdk new performance-operator --repo github.com/openshift-kni/performance-addon-operators --vendor=true
The PerformanceProfile
CRD is the API of the performance operator and offers high level options
for applying various performance tunings to cluster nodes. The API and its controller were created with:
operator-sdk add api --api-version=performance.openshift.io/v1alpha1 --kind=PerformanceProfile
operator-sdk add controller --api-version=performance.openshift.io/v1alpha1 --kind=PerformanceProfile
The performance profile API is documented in detail in the Performance Profile doc. Follow the API versions doc to check the supported API versions.
Developers can build and push the Performance Operator images from the source tree using make targets.
export REGISTRY_NAMESPACE=<your quay.io namespace>
export IMAGE_TAG=<the image tag to use> #defaults to "latest"
make build-containers
make push-containers
The building of the index image requires that the bundle image will be public available under the image registry, otherwise the creation of the index image will fail.
It is number of ENV variables that you should define to get all components to have the right z-stream version 4.y.z
and allow the seamless upgrade.
export CSV_CHANNEL="4.y"
export CSV_VERSION="4.y.z"
export CSV_FROM_VERSION="4.y.z-1"
export IMAGE_TAG=4.x.y
export OPERATOR_VERSION=4.y.z
export CSV_SKIP_RANGE=">=4.y-1.0 <4.y.z
make build-containers
make push-containers
If you use your own images, make sure they are made public in your quay.io account!
If you want to use the performance operator's upstream images,
- unset
REGISTRY_NAMESPACE
(it will default toopenshift-kni
) - if you deploy on OCP 4.4, run
export IMAGE_TAG=v4.4
- if you deploy on OCP 4.5, unset
IMAGE_TAG
(it will default tolatest
)
Deploy the operator by running:
CLUSTER=manual make cluster-deploy
This will deploy
- a
MachineConfigPool
for the nodes which will be tuned - all manifests for letting OCP's
Operator Lifecycle Manager (OLM)
deploy the Performance Operator:- a
CatalogSource
- a
Namespace
- a
OperatorGroup
- a
Subscription
- a
- a
PerformanceProfile
Note: The performance operator supports only AllNamespaces
InstallModeType for OperatorGroup
.
The deployment will be retried in a loop until everything is deployed successfully, or until it times out.
Note:
CLUSTER=manual
lets the deploy script use thecluster-setup/manual-cluster/performance/
kustomization directory. In CI thecluster-setup/ci-cluster/performance/
dir will be used. The difference is that the CI cluster will deploy the PerformanceProfile in the test code, while themanual
cluster includes it in the kustomize based deployment.
Now you need to label the nodes which should be tuned. This can be done with
make cluster-label-worker-cnf
This will label 1 worker node with the worker-cnf
role, and OCP's Machine Config Operator
will start tuning this node.
In order to wait until MCO is ready, you can watch the MachineConfigPool
until it is marked as updated with
CLUSTER=manual make cluster-wait-for-mcp
Note: Be aware this can take quite a while (many minutes)
Note: in CI this step is skipped, because the test code will wait for the MCP being up to date.
The operator can render manifests for all the components it supposes to create, based on Given a PerformanceProfile
You need to provide the following environment variables
export PERFORMANCE_PROFILE_INPUT_FILES=<your PerformanceProfile directory path>
export ASSET_OUTPUT_DIR=<output path for the rendered manifests>
Build and invoke the binary
build/_output/bin/performance-addon-operators render
Or provide the variables via command line arguments
build/_output/bin/performance-addon-operators render --performance-profile-input-files <path> --asset-output-dir<path>
When the deployment fails, or the performance tuning does not work as expected, follow the Troubleshooting Guide for debugging the cluster. Please provide as much info from troubleshooting as possible when reporting issues. Thanks!
Unit tests can be executed with make unittests
.
The functional tests are located in /functests
. They can be executed with make functests-only
on a cluster with a
deployed Performance Operator and configured MCP and nodes. It will create its own Performance profile!
The latency-test container image gives the possibility to run the latency test without need to install go, ginkgo or other go related modules.
The test itself is running the oslat
cyclictest
and hwlatdetect
binaries and verifies if the maximal latency returned by each one of the tools is
less than specified value under the MAXIMUM_LATENCY
.
To run the latency test inside the container:
docker run --rm -v /kubeconfig:/kubeconfig \
-e KUBECONFIG=/kubeconfig \
-e LATENCY_TEST_RUN=true \
-e LATENCY_TEST_RUNTIME=60 \
-e MAXIMUM_LATENCY=700 \
quay.io/openshift-kni/cnf-tests /usr/bin/run-tests.sh
You can run the container with different ENV variables, but the bare minimum is to pass
KUBECONFIG
mount and ENV variable, to give to the test access to the cluster and
LATENCY_TEST_RUN=true
to run the latency test.
LATENCY_TEST_DELAY
indicates an (optional) delay in seconds to be used between the container is created and the tests actually start. Default is zero (start immediately).LATENCY_TEST_RUN
indicates if the latency test should run.LATENCY_TEST_RUNTIME
the amount of time in seconds that the latency test should run.LATENCY_TEST_IMAGE
the image that used under the latency test.LATECNY_TEST_CPUS
the amount of CPUs the pod which run the latency test should requestOSLAT_MAXIMUM_LATENCY
the expected maximum latency for all buckets in us in the oslat test.CYCLICTEST_MAXIMUM_LATENCY
the expected maximum latency for the cyclictest test.HWLATDETECT_MAXIMUM_LATENCY
the expected maximum latency for the hwlatdetect test.MAXIMUM_LATENCY
a unified value for the expected maximum latency for all tests (In case both provided, the specific variables will have precedence over the unified one).
See CONTRIBUTING for some guidelines.
A custom CSV entry for the registry container can be generated using the
make generate-csv
make target.
First export your CSV details as environment variables.
required options
export IMAGE_REGISTRY="quay.io"
export REGISTRY_NAMESPACE="some-operator=repo-namespace"
export IMAGE_TAG="some-operator-image-tag"
export CSV_VERSION="0.0.3"
optional options
export REPLACES_CSV_VERSION="0.0.2"
export CSV_SKIP_RANGE=">=0.0.1 <0.0.2"
Then run make generate-csv
The result will be stored in the deploy/olm-catalog/performance-addon-operator
directory within a directory that matches the CSV_VERSION
set.
Running make registry-container
after creating a new custom CSV will result
in a registry bundle that includes the new CSV version and all other CSV
versions in the deploy/olm-catalog
directory.