Network Bandwidth-Aware Kubernetes Cluster
summary
Prototype cluster-level bandwidth reservation and placement system that adds a scheduler extender plus a node agent to register per-node ingress/egress capacity and per-pod requests, uses node labels and Cluster API to expose capacities, and enforces queues via Linux tc and OVS. Scheduler filters out nodes with insufficient reserved bandwidth and scores by remaining capacity (example cluster: three 1000 Mbps nodes and one 300 Mbps node; the 300 Mbps node is filtered). Prototype works to reduce contention but admits reservations are compressible so not a strict hard guarantee.
AI-written summary · about
chapters (8)
- 0:00Start
- 6:40Network Features in Kubernetes (2/3)
- 13:23BareMetal Deployment Physical NIC
- 15:18Virtual Machine Deployment Kubernetes Cluster on vSphere with NSXT
- 17:10Workflow
- 19:45Kubernetes Cluster API
- 21:45Mark Pods with Resource Request
- 24:31Algorithm for Extended Rules
related talks
36mNetwork Bandwidth-Aware Kubernetes Cluster
36mNetwork-aware Scheduling in Kubernetes
30mSched.net: A Network-Aware Kubernetes Scheduler [I]
32mCoordinate Workloads Colocation: QoS-Oriented Scheduling Enhancement on K8s
30mEnhancing the Kubernetes Scheduler for Diverse Workloads in Large Clusters
34mNetwork QoS Support for Kubernetes Applications
38mDynamic Pod Resource Boundary Adjustment in Web Scale Clusters
18mLet There Be Topology-Awareness in Kube-Scheduler!