NHD - A Topology-Aware Scheduler for K8s for Low-Latency & HPC Applications
Talkat KubeCon + CloudNativeCon North America 2019 · on Wednesday, 2019-11-20· at 14:25· Pacific Ballroom, Salon 14-15 - Marriott Marquis San Diego Marina Hotel·
summary
A Kubernetes custom scheduler that places pods by NUMA + PCIe + hyperthreading topology and enforces core isolation via an isolcpus + cpusets daemon, it schedules DPDK-style, low-latency workloads (targeting sub-10µs packet processing) by matching requested hugepages, GPUs and SR‑IOV virtual functions to node topology, writing physical IDs into pod config maps. Implemented as a Python3 gRPC scheduler, it binds pods through the API, is open source, runs on ~30 nodes, and is slower than upstream scheduler but functional.
AI-written summary · about
chapters (40)
- 0:00Introduction
- 0:25Agenda
- 0:52Room Numbers
- 1:12Terminology
- 2:05Interfaces
- 3:04User Space Networking
- 5:57TopologyAware Scheduler
- 6:39Single Socket Scheduler
- 10:14PCIe Topology
- 11:52Challenges
- 12:05Isolating CPUs
- 14:02CPU Sets
- 15:53ISIL CPU Set
- 16:48CPU Selection
- 18:04NIC Sharing
- 18:46Virtual Switch
- 20:02Host Device
- 20:37SRIOV
- 21:47Topology Aware Scheduler
- 22:48Config Translation
- 23:27Matcher
- 24:32Data Plane
- 26:36Hyperthreading
- 28:06Binding Pod to Node
- 28:34Topology Manager
- 29:34Future Work
- 32:00Thanks
- 32:35Nvidia multiprocess
- 34:05Config map
- 34:50Node feature discovery
- 35:50Custom schedulers
- 37:23Why custom schedulers
- 38:22Are we using ISIL CPUs
- 39:49How do you specify the CPU cores
- 40:40Why do you use a config map
- 41:24Will it work on PowerPC
- 42:22Why did you change the host pet
- 42:59Nvidia Device Plugin
- 45:05GPU Health Check
- 45:44Special Privileges
related talks
18mLet There Be Topology-Awareness in Kube-Scheduler!
34mUnlocking Performance: Topology-Aware CPU Scheduling W...
36mNetwork-aware Scheduling in Kubernetes
24mAccelerate Your AI/ML Workloads With Topology-Aware Scheduling in Kueue
30mSched.net: A Network-Aware Kubernetes Scheduler [I]
20mBoosting Training and Inference Performance via Topology-Aware Scheduling of Heterogeneous Resources
10mFair Scheduling for Deep Learning Workloads in Kubernetes
31mSIG Scheduling Update: Transition From Pod To Workload Scheduling