Skip to content

hygon-dcu.md

Latest commit

 

History

History
93 lines (70 loc) · 3.24 KB

hygon-dcu.md

File metadata and controls

93 lines (70 loc) · 3.24 KB
Raw

Hygon DCU

Deploy HAMi-DRA webhook for Hygon DCU clusters that already run k8s-dcu-dra-driver.

This chart deploys only the mutating/validating webhook and TLS certificates. The DCU DRA driver and DeviceClass (dra.hygon.com) must be deployed separately via k8s-dcu-dra-driver.

Prerequisites

  1. Kubernetes with DRA enabled (same requirements as the main chart).
  2. k8s-dcu-dra-driver deployed; DeviceClass dra.hygon.com must exist.
  3. cert-manager installed.
  4. hami-dra-webhook image built and reachable from all nodes (override webhook.image.* if not using the default registry).

Install

From a local checkout:

helm install hami-dra ./charts/hami-dra \
  -n hami-system --create-namespace \
  --set deviceVendor=hygon \
  --set drivers.nvidia.enabled=false \
  --set monitor.enabled=false

From the published chart:

helm install hami-dra hami-dra/hami-dra \
  -n hami-system --create-namespace \
  --set deviceVendor=hygon \
  --set drivers.nvidia.enabled=false \
  --set monitor.enabled=false

Upgrade with the same flags:

helm upgrade hami-dra ./charts/hami-dra -n hami-system \
  --set deviceVendor=hygon \
  --set drivers.nvidia.enabled=false \
  --set monitor.enabled=false

Important values

Value Recommended Notes
deviceVendor hygon Switches webhook to Hygon DCU resource names and dra.hygon.com driver.
drivers.nvidia.enabled false Do not deploy the NVIDIA DRA driver DaemonSet.
drivers.dcu.deviceClassName / driverName empty (default) Optional overrides for webhook config; falls back to dcuDeviceClassName / dcuDraDriverName. Does not deploy a driver.
monitor.enabled false Monitor is NVIDIA-oriented today; disable for DCU-only clusters.
certs.certManager.enabled true (default) Uses cert-manager for webhook TLS.

DCU resource names and driver identifiers use chart defaults in values.yaml (no override needed in most cases):

dcuResourceName: "hygon.com/dcunum"
dcuResourceMem: "hygon.com/dcumem"
dcuResourceCores: "hygon.com/dcucores"
dcuDeviceClassName: dra.hygon.com
dcuDraDriverName: dra.hygon.com

Fractional compute (hygon.com/dcucores)

When Pods request compute as a percentage via hygon.com/dcucores, set dcuReferenceComputeUnits to one card's total compute units from the cluster:

kubectl get resourceslice -o yaml

Example:

helm upgrade hami-dra ./charts/hami-dra -n hami-system \
  --set deviceVendor=hygon \
  --set drivers.nvidia.enabled=false \
  --set monitor.enabled=false \
  --set dcuReferenceComputeUnits=128

For whole-card requests only (hygon.com/dcunum without dcumem / dcucores), leave dcuReferenceComputeUnits at 0 (default).

What gets deployed

Component Hygon DCU install Default NVIDIA install
hami-dra-webhook Yes (Hygon vendor) Yes (NVIDIA vendor)
cert-manager resources Yes Yes
NVIDIA DRA driver DaemonSet No Yes
hami-dra-monitor No Yes
DCU DRA driver External (k8s-dcu-dra-driver) N/A