WG Device Management Charter

Mon, 01 Jan 0001 00:00:00 +0000

WG Device Management Charter

This charter adheres to the conventions described in the Kubernetes Charter README and uses the Roles and Organization Management outlined in wg-governance .

Scope

Enable simple and efficient configuration, sharing, and allocation of accelerators and other specialized devices. This working group focuses on the APIs, abstractions, and feature designs needed to configure, target, and share the necessary hardware for both batch and serving (inference) workloads.

In scope

Enable efficient utilization of specialized hardware devices. This includes sharing one or more resources effectively (many workloads sharing a pool of devices), as well as sharing individual devices effectively (several workloads dividing up a single device for sharing).
Enable workload authors to specify “just enough” details about their workload requirements to ensure it runs optimally, without having to understand exactly how the infrastructure team has provisioned the cluster.
Enable the scheduler to choose the correct place to run a workload the vast majority of the time (rejections should be extremely rare).
Enable cluster autoscalers and other node auto-provisioning components to predict whether creating additional resources will satisfy workload needs, before provisioning those resources.
Enable the shift from “pods run on nodes” to “workloads consume capacity”. This allows Kubernetes to provision sets of pods on top of sets of nodes and specialized hardware, while taking into account the relationships between those infrastructure components.
Enable in-node devices as well as network-accessible devices.
Minimize workload disruption due to hardware failures.
Address fragmentation of accelerator due to fractional use.
Additional problems that may be identified and deemed in scope as we gather use cases and requirements from WG Serving, WG Batch, and other stakeholders.
Address all of the above while with a simple API that is a natural extension of the existing Kubernetes APIs, and avoids or minimizes any transition effort.

Out of Scope

Higher-level workload controller APIs (for example, the equivalent of Deployment, StatefulSet, or DaemonSet) for specific types of workloads.
General resource management requirements not related to devices.

Deliverables

The WG will coordinate the delivery of KEPs and their implementations by the participating SIGs. Interim artifacts will include documents capturing use cases, requirements, and designs; however, all of those will eventually result in KEPs and code owned by SIGs.

WG Device Management on Kubernetes Contributors

WG Device Management Charter

WG Device Management Charter

Scope

In scope

Out of Scope

Deliverables