px-bench

A framework for running storage performance benchmarks on Kubernetes based platforms.

• git clone the repository to a Linux (macOS may work but not tested) system with kubectl access to the K8s cluster.
• If necessary, build the image and push to your registry:

cd image
docker build -t px-bench .
docker push ...

• TBD: Create a wrapper script that will create the namespace and apply all YAML in order and with correct timing!
• Create the px-bench namespace with kubectl create ns px-bench
• Edit px-bench-env.yml to set the ConfigMap env to set desired values. If necessary, update image: to reflect the image you built. NOTE: SET YOUR STORAGECLASSES IN THIS FILE!
• Do NOT edit px-bench-main.yml (unless you are attempting to change the behavior of the benchmark!)
• In order to consume most of the available RAM so it is not used for buffering, run kubectl apply -f chewram.yml.
• Wait for kubectl -n px-bench get pod -n chewram for all the pods to show as 1/1 Running.
• kubectl apply -f px-bench-env.yml to apply the configuration settings.
• kubectl apply -f px-bench-main.yml to start the run (ignore the runAsNonRoot warning if on Openshift).
• Monitor progress with kubectl logs -n px-bench -l px-bench=fio -f. With the defaults, runtime is expected to be around 15 minutes.
• Wait for kubectl get pod -n px-bench for all the pods to show as Completed.

This will iterate through the combinations of blocksize_list, readwrite_list, and storageclass_list set in the ConfigMap, and runs those as independent fio jobs. Output will go the ConfigMap fio-output. Configurations will go to the ConfigMap fio-config.

To retrieve the output for processing, run kubectl get cm csv -n px-bench -o jsonpath='{.data.csv}'. Open the spreadsheet and make a copy. Paste the CSV into the sheet at A1 and make sure "Split text to columns" is selected.

Click the Extensions menu and select "Apps Script". Click Run. This should create a number of new sheets populated with some bar charts.

When complete, to clean up, you can either remove the jobs, pods and PVCs from the namespace, or delete the namespace entirely.