Release Notes

Change Description

Kunpeng BoostKit addresses the challenge of static CPU management policy limitations in large-scale container deployments with Kunpeng Topology Affinity Plugin (TAP), a Kubernetes NUMA affinity scheduling plugin. This plugin dynamically optimizes pod CPU allocation during node deployment by intelligently adjusting scheduling scopes according to real-time compute node load conditions, effectively maintaining CPU affinity. Currently, this feature is incompatible with the native Kubernetes topology manager.

Supported functions:

When the number of CPUs in a NUMA node of a worker node reaches the CPU limit declared by the container, the plugin can restrict the scheduling range of all CPU time slices of the container to this NUMA node.

New features:

• The newly introduced topology_aware allocation policy enables hierarchical resource management by leveraging system CPU and memory topology. This intelligent approach binds CPU and memory resources for burstable-level pods and containers across multiple layers including socket, die, NUMA node, and core while prioritizing allocation for GPUs within specific NUMA nodes.
• Expanded runtime compatibility now supports containerd environments by intercepting CRI container requests to optimize resource distribution.

Resolved Issues

None.

Known Issues

None.