Elastic Kubernetes Service (VKS)
Managed Kubernetes for GPU workloads — GPU labels, RDMA, persistent volumes, service publishing
Virtual Kubernetes Service (VKS) is the managed Kubernetes offering on Alaya NeW Cloud, with first-class support for GPU scheduling, cross-node RDMA, shared / dynamic storage, and service publishing — so training and inference workloads consume GPU on demand.
Capabilities
- GPU resources — H800 / L40S and more, requested by label, with CPU and memory pre-sized to GPU count.
- Cross-node RDMA — IB / RoCE built in, enabled with one line
rdma/rdma_shared_device_ain your YAML. - Persistent storage — shared static PVCs across accounts plus on-demand dynamic PVCs; underlying NAS / block details are abstracted away.
- Service publishing — built-in
ServiceExporterauto-mints a public domain; no separate load balancer needed. - Multi-tenancy — primary-account-driven authorization, with per-user parallelism and consumption rules on dedicated clusters.
Learning path
Concepts
Pod, Workload, Service, PVC, Namespace — quick reference
Authorize VKS
Primary account grants access to team members
Use the cluster
Download kubeconfig and connect in 5 minutes
Cluster detail
Resource management, scaling, parallelism and consumption rules
GPU resources
GPU model table and YAML request examples
Cross-node RDMA
Enable high-performance networking inside Pods
Create a PersistentVolumeClaim
Static and dynamic PVC declarations and mounts
Create a container
Pod and Deployment templates
Create a service
Expose and access workloads inside the cluster
Publish a service
Push services to the public network with ServiceExporter
Advanced recipes
Common extensions on top of VKS — full guides live in the Cookbook:
-
Remote development SSH · VS Code in browser · Jupyter · Dev Containers · DevPod
-
HPC scheduling Slurm on VKS
-
Package management Helm essentials
VKS vs CCI
CCI is for single-container ad-hoc workloads. VKS is for multi-container, multi-service orchestration. They share the same underlying GPU pool and can be combined.
Last updated on
