Performance Testing

Contributors kevin-hoke mboglesby Download PDF of this page

We performed a simple performance comparison as part of the creation of this solution. We executed several standard NetApp AI benchmarking jobs by using Kubernetes, and we compared the benchmark results with executions that were performed by using a simple Docker run command. We did not see any noticeable differences in performance. Therefore, we concluded that the use of Kubernetes to orchestrate containerized AI training jobs does not adversely affect performance. See the following table for the results of our performance comparison.

Benchmark Dataset Docker Run (images/sec) Kubernetes (images/sec)

Single-node TensorFlow

Synthetic data

6,667.2475

6,661.93125

Single-node TensorFlow

ImageNet

6,570.2025

6,530.59125

Synchronous distributed two-node TensorFlow

Synthetic data

13,213.70625

13,218.288125

Synchronous distributed two-node TensorFlow

ImageNet

12,941.69125

12,881.33875