PKS is designed to work with vSphere for on-premises deployments. Kubernetes worker nodes are deployed, managed, and scaled as virtual machines. PKS supports deploying single, large, distributed Kubernetes clusters and many small clusters. VMware NSX provides load balancing and other networking services such as micro segmentation. Users can choose to deploy PKS on a VMware VSAN-enabled cluster or use an external storage platform such as Isilon to provide blob storage to Domino Data Science Platform.
The following figure shows the PKS architecture:
Figure 11. Pivotal Container Service
GPU-enabled resources can be made available to PKS Kubernetes clusters through VMware’s Bitfusion software. Bitfusion software enables dynamic GPU allocation over RDMA-enabled networks. A breakout session at VMworld 2019 presented the technology behind Bitfusion and demonstrated how to connect GPU and FPGA resources over network fabric. For more information, see this VMware vSphere Blog. We will update this white paper when support for these technologies are available in PKS Enterprise.
For more information, about the Pivotal Ready Architecture from Dell EMC, see the Pivotal Ready Architecture Datasheet.
During the evaluation of Domino Data Science Platform, we compared using the application on both open-source and PKS distributions of Kubernetes. The following table describes our experience with these two distributions:
Table 8. Open-source Kubernetes and PKS Enterprise comparison
|
Open-source Kubernetes |
PKS Enterprise |
Deployment |
Complex – DIY: Customers are responsible for deploying and managing their Kubernetes environment. It is the responsibility of the customer to select the correct Kubernetes components. |
Turnkey: Through PKS enterprise, production grade Kubernetes clusters can be deployed on a virtual infrastructure with advanced networking, a built-in private registry with enterprise security features, and full life cycle management support of the clusters. |
Enterprise support |
Not available |
Yes. Provided by VMware. Delivers a stable Kubernetes build. |
Training and inference using CPU resources |
Yes |
Yes |
Training and inference using GPU resources |
Yes |
Not available |