Benchmarking Kubernetes Persistent Volumes with fio: IOPS, Bandwidth, and Latency

Tool

Dbench

GitHub repository: https://github.com/leeliu/dbench

Usage

Edit the dbench.yaml file to set the storageClassName that matches your own Storage Class. kubectl get storageclasses Deploy the job: kubectl apply -f dbench.yaml After deployment, the Dbench Job will:

Use storageClassName: ssd (default) to provision a 1000Gi persistent volume (default).

Run a series of fio tests on the newly provisioned disk.

Execute 9 tests, each lasting 15 seconds, for a total runtime of about 2.5 minutes.

Track benchmark progress with: kubectl logs -f job/dbench An empty output means the job has not been created yet or the storageClassName is invalid; see the troubleshooting section.

When all tests finish, a summary similar to the following appears:

==================
= Dbench Summary =
==================
Random Read/Write IOPS: 75.7k/59.7k. BW: 523MiB/s / 500MiB/s
Average Latency (usec) Read/Write: 183.07/76.91
Sequential Read/Write: 536MiB/s / 512MiB/s
Mixed Random Read/Write IOPS: 43.1k/14.4k

The summary reports Random Read/Write IOPS, bandwidth (BW), average latency, sequential throughput, and mixed random IOPS.

Clean up after testing:

kubectl delete -f dbench.yaml

Notes / Troubleshooting

If a Persistent Volume Claim stays Pending, you likely did not specify a valid Storage Class. Verify with kubectl get storageclasses and ensure the requested size is 1000Gi (default).

Binding the volume may take time; the Kubernetes Dashboard UI may show the Dbench Job in red until the volume is ready.

Testing different disk sizes is useful because many cloud providers price IOPS per GB. A 4000Gi volume may deliver higher performance than a 1000Gi volume. To retest, edit the YAML, delete the job, wait for deprovisioning, then re‑apply.

All fio test definitions reside in docker-entrypoint.sh, which runs the following steps:

Testing Read IOPS...

Testing Write IOPS...

Testing Read Bandwidth...

Testing Write Bandwidth...

Testing Read Latency...

Testing Write Latency...

Testing Read Sequential Speed...

Testing Write Sequential Speed...

Testing Read/Write Mixed...

Tencent Cloud K8S Cluster Production Practice

List storage classes:

kubectl get storageclass

Edit dbench.yaml :

vi dbench.yaml

Apply the job: kubectl apply -f dbench.yaml Watch logs:

kubectl logs -f job/dbench

Delete the job when finished:

kubectl delete -f dbench.yaml