Containerized workloads and microservices are designed to be ephemeral. It can be to answer the needs of scalability (via Pod autoscaling) or because they process data for a short period (ML workload). Every day stateful apps are becoming a bigger part of that Kubernetes workloads.
In that context, it is vital for the storage backend to be able to serve and delete volumes & snapshots as fast as possible with a low impact on the performance of the worker nodes.
TODO : need PowerFlex SME for more accurate technical explanations & Flesh out
The CSI driver for PowerFlex is capable of blazing-fast volume provisioning thanks to its software first architecture. The PowerFlex Software Data Client directly initiates a network connection to the Storage Data Servers and mounts the block device. This network-first approach bypass long steps like SCSI Bus rescan and leads to saving precious time in a rapidly changing ecosystem.
To measure the performance of the CSI driver, we can measure the time between Kubernetes events ; for example, the time it takes between a volume create request and volume bounded, or Pod created and started, etc.
This benchmark has been done with vanilla Kubernetes StatefulSet and the help of a kube-event-export/elasticsearch/kibana stack for the reporting.
The performance of the CSI driver for PowerFlex has a steady performance with low impact on CPU performance.
For example, it takes less than 30 seconds to provision 100 volumes in PowerFlex and make them available to Kubernetes It takes an extra 3 minutes to format and mount the volumes them to a single Pod.
On another run, it takes 7 minutes seconds from start to finish to provision 100 Pods with 4 volumes each. In this test, the Pods are scheduled on three different nodes.
With pre-provisioned and pre-formatted volumes for the same configuration (100 Pods with 4 volumes each), it takes about the same time.
During all that time the CPU usage on the controller pod, the impact was limited:
