Enhancement Description
Using DRA drivers for core resources (like dra-driver-cpu), enables fine-grained resource management but introduces a critical accounting problem. The standard kube-scheduler resource accounting model is unaware of allocations made by the DynamicResources plugin, and vice-versa, leading to a potential resource oversubscription on nodes. While a synchronization model similar to the one used for DRA Extended Resources is needed, the same solution cannot be applied directly. The key difference is that extended resources are advertised either through the node object (node.status.allocatable) or via ResourceSlice, but not both simultaneously for the same node. In contrast, core resources like CPU are always present in node.status.allocatable, and a DRA driver would also manage them via ResourceSlice. This simultaneous representation creates a unique challenge that requires a unified accounting model.
cc: @johnbelamaric, @mortent , @ffromani , @yujuhong , @dchen1107
Please keep this description up to date. This will help the Enhancement Team to track the evolution of the enhancement efficiently.
Enhancement Description
Using DRA drivers for core resources (like dra-driver-cpu), enables fine-grained resource management but introduces a critical accounting problem. The standard kube-scheduler resource accounting model is unaware of allocations made by the
DynamicResourcesplugin, and vice-versa, leading to a potential resource oversubscription on nodes. While a synchronization model similar to the one used for DRA Extended Resources is needed, the same solution cannot be applied directly. The key difference is that extended resources are advertised either through the node object (node.status.allocatable) or viaResourceSlice, but not both simultaneously for the same node. In contrast, core resources like CPU are always present innode.status.allocatable, and a DRA driver would also manage them viaResourceSlice. This simultaneous representation creates a unique challenge that requires a unified accounting model.cc: @johnbelamaric, @mortent , @ffromani , @yujuhong , @dchen1107
k/enhancements) update PR(s): KEP-5517: Introduce DRA for Node Allocatable Resources KEP #5755k/k) update PR(s): Introduce support of DRA for Node Allocatable Resources kubernetes#136725k/website) update PR(s): Docs update for KEP-5517: DRA for Node Allocatable Resources website#54598k/enhancements) update PR(s): KEP-5517: DRA Node Allocatable Resources alpha2 changes #6082k/k) update PR(s): KEP-5517: DRANodeAllocatableResources alpha 2 implementation kubernetes#140009k/website) update(s): Docs update for KEP-5517 alpha2 website#56318k/enhancements) update PR(s): KEP-5517: DRA Node Allocatable Resources beta #6324k/k) update PR(s):k/website) update(s):Please keep this description up to date. This will help the Enhancement Team to track the evolution of the enhancement efficiently.