Skip to main content
Version: Next

Enable Cambricon MLU Sharing

HAMi now supports cambricon.com/mlu by implementing most device-sharing features similar to NVIDIA GPUs, including:

  • MLU Sharing: Tasks can request a fraction of an MLU instead of an entire MLU card. This enables multiple tasks to share the same MLU device.

  • Device Memory Control: You can allocate MLUs with a specified memory size, with guaranteed enforcement to ensure usage does not exceed the requested limit.

  • Device Core Control: MLUs can be assigned a specific number of compute cores, and enforcement ensures core usage remains within bounds.

  • MLU Type Selection: You can use annotations to specify which MLU types a task must use or must avoid by setting cambricon.com/use-mlutype or cambricon.com/nouse-mlutype.

Prerequisites​

  • neuware-mlu370-driver > 5.10
  • cntoolkit > 2.5.3

Enabling MLU Sharing​

  1. Install HAMi via Helm

    Follow the instructions under the Enabling vGPU Support in Kubernetes section in the HAMi README.

  2. Enable SMLU mode on each MLU device

    cnmon set -c 0 -smlu on
    cnmon set -c 1 -smlu on
    # Repeat for all devices...

  3. Deploy the Cambricon device plugin

    Get the cambricon-device-plugin from your device provider, and configure it with the following parameters:

    • mode=dynamic-smlu: Enables dynamic SMLU support.
    • min-dsmlu-unit=256: Sets the minimum allocatable memory unit to 256 MB.

    Refer to your provider’s documentation for additional details.

  4. Apply the configured plugin

    kubectl apply -f cambricon-device-plugin-daemonset.yaml

Running MLU Jobs​

To request shared MLU resources in a container, use the following resource types:

  • cambricon.com/vmlu
  • cambricon.com/mlu.smlu.vmemory
  • cambricon.com/mlu.smlu.vcore

Here is an YAML example:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: binpack-1
  labels:
    app: binpack-1
spec:
  replicas: 1
  selector:
    matchLabels:
      app: binpack-1
  template:
    metadata:
      labels:
        app: binpack-1
    spec:
      containers:
        - name: c-1
          image: ubuntu:18.04
          command: ["sleep"]
          args: ["100000"]
          resources:
            limits:
              cambricon.com/vmlu: "1"
              cambricon.com/mlu.smlu.vmemory: "20"
              cambricon.com/mlu.smlu.vcore: "10"
note

  1. Init containers are not supported for MLU sharing.

    Pods with the cambricon.com/mlumem resource specified in an init container will not be scheduled.

  2. Resource constraints only apply to shared mode (vmlu=1).

    The cambricon.com/mlu.smlu.vmemory and cambricon.com/mlu.smlu.vcore resources are only effective when cambricon.com/vmlu is set to 1. If vmlu > 1, a full MLU device will be allocated regardless of vmemory and vcore values.