DeviceShare

Introduction

The DeviceShare Plugin is an advanced resource scheduling plugin in Volcano that provides a common framework for sharing specialized hardware devices (like GPUs, NPUs, FPGAs) across multiple pods.

Rather than implementing fragmented logic for each new hardware accelerator, Volcano exposes a unified Devices interface. The deviceshare plugin leverages this interface to perform robust allocation, node filtering, and resource tracking for shared devices.

Mechanism

The DeviceShare plugin works in conjunction with device-specific implementations. It exposes standard scheduling operations such as Predicate (filtering nodes based on available device capacity) and Allocate/Release (assigning portions of a device to specific pods).

Currently, the deviceshare plugin serves as the underlying engine powering features like:

• GPU Sharing: Allowing multiple pods to request fractions of a single physical GPU's memory.
• vGPU and vNPU: Virtualizing accelerator slices.
• GPU Exclusive: Restricting a pod to exclusively own a GPU to avoid contention.

Configuration and Usage

The deviceshare plugin is typically enabled implicitly when you enable device sharing predicates in the Volcano scheduler config map. However, if you are developing custom device sharing logic or need to explicitly declare it, it can be configured in your volcano-scheduler-configmap:

actions: "enqueue, allocate, backfill"
tiers:
  - plugins:
      - name: priority
      - name: gang
      - name: conformance
      - name: deviceshare   # Enable the device share framework plugin
        arguments:
          deviceshare.GPUSharingEnable: true
          deviceshare.VGPUEnable: true
  - plugins:
      - name: overcommit
      - name: drf
      - name: predicates
      - name: proportion
      - name: nodeorder
      - name: binpack

Arguments

The deviceshare plugin supports several arguments to configure specific hardware accelerator behaviors and scheduling policies:

Argument	Type	Default	Description
deviceshare.GPUSharingEnable	bool	false	Enables GPU sharing, allowing multiple pods to request fractions of a single physical GPU.
deviceshare.GPUNumberEnable	bool	false	Enables GPU number scheduling, allowing pods to request a specific number of exclusive GPUs.
deviceshare.VGPUEnable	bool	false	Enables vGPU (Virtual GPU) scheduling.
deviceshare.AscendHAMiVNPUEnable	bool	false	Enables Ascend vNPU scheduling via the HAMi (Heterogeneous AI Computing Virtualization Middleware) mode.
deviceshare.AscendMindClusterVNPUEnable	bool	false	Enables Ascend vNPU scheduling via the official MindCluster mode.
deviceshare.SchedulePolicy	string	binpack	Defines the node scoring policy for shared devices. Supports binpack (pack workloads densely onto devices) and spread (distribute workloads evenly across devices).
deviceshare.ScheduleWeight	int	1	The weight multiplier applied to the SchedulePolicy score during the NodeOrder scoring phase.
deviceshare.KnownGeometriesCMName	string	hami-scheduler-device	Specifies the ConfigMap name used to store known device geometries and topologies.
deviceshare.KnownGeometriesCMNamespace	string	kube-system	Specifies the namespace where the KnownGeometries ConfigMap is located.

Note: For specific guides on how to configure your workloads to request shared GPUs or NPUs, please refer to the dedicated guides for GPU Sharing and vNPU. For vGPU and MIG configurations, see GPU Virtualization.