What’s in the Box? Resource Management in Kubernetes
summary
Concrete extensions to Kubernetes are described: CPU manager adds static CPU pinning and thread-level pinning (now in upstream with static policy), HugePages becomes a first-class preallocated resource (2M/1G) and an emptyDir type, and device plugins let vendors advertise and allocate hardware without forking kubelet (alpha in 1.8, beta by 1.10). A proposed new manager aligns NUMA allocations across CPU and devices. Demoed packet-rate gains up to 3.9x for a DPDK workload and published container bare-metal experience kits.
AI-written summary · about
chapters (10)
- 0:00Start
- 1:24KUBERNETES RESOURCE MANAGEMENT TODAY
- 3:32ADDRESS KEY CHALLENGES IN CONTAINERS BARE METAL
- 6:04NODE FEATURE DISCOVERY
- 8:27CPU MANAGER FOR KUBERNETES - CPU PINNING AND ISOLATION
- 12:23HUGE PAGE NATIVE SUPPORT IN KUBERNETES
- 13:35DEVICE PLUGINS OVERVIEW
- 14:56QAT SUPPORT IN KUBERNETES
- 16:35NUMA MANAGER FOR KUBERNETES - NUMA ALIGNMENT OF RESOURCES
- 18:48CONTAINER BARE METAL EXPERIENCE KITS
related talks
36mTaming the Beast: Advanced Resource Management With Kubernetes
41mPanel Discussion: State of Art for Enabling Performance Sensitive Workloads and What We Need Next
21mDemystifying Resource Management in Kubernetes
34mUnlocking Performance: Topology-Aware CPU Scheduling W...
21mDo Not Disturb Mode: Fine-Tuning Resources for Latency Sensitive Workloads
18mLet There Be Topology-Awareness in Kube-Scheduler!
36mMitigating Noisy Neighbours: Advanced Container Resource Management
29mRight-Size Your Kubernetes Deployments