Five Ways With a CNI: Understanding Kubernetes Networking For Performance..
Talkat KubeCon + CloudNativeCon North America 2022 · on Thursday, 2022-10-27· at 14:30· Virtual Platform Only·
summary
Lab uses HP DL380 Gen10 nodes with ConnectX‑6 2×100Gb, Ubuntu 20.04, Kubernetes 1.23.7, iperf2, Intel IMB, Kubeflow MPI operator, Volcano scheduler, OpenFOAM and an Oxford Nanopore basecaller. Calico matches host TCP bandwidth and costs ~30% system CPU, OVN without offload drops sharply under contention, SR‑IOV hits a 100Gb/s port limit but gives SR-IOV latency ~1.83µs, and hardware offloaded OVN with RDMA both cuts CPU to ~15% and achieves accelerated OVN saturates 200Gb/s and ~180Gb/s for 32MB MPI, halving OpenFOAM runtimes.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:33High Performance Computing (HPC)
- 1:09Artificial Intelligence (AI)
- 1:44AI & HPC - Network Considerations
- 3:27Remote Direct Memory Access - RDMA
- 6:20OVN Kubernetes
- 7:22Single Root I/O Virtualization (SR-IOV)
- 9:06Legacy OVS / OVN Challenges
- 12:29Lab Set-Up: K8s Bare Metal
- 14:14Lab Set-Up: K8s Over OpenStack VM
- 16:03Test Scenarios
- 16:54Iperf 2 - TCP Throughput Benchmark
- 24:36Calico vs Acc OVN: CPU load during test
- 25:30Accelerated OVN: Flow offloads during test
- 26:04Intel MPI Benchmarks - PingPong
- 28:42Computational Fluid Dynamics - OpenFOAM
- 32:01Genome Sequencing - Basecalling benchmark
related talks
30mSched.net: A Network-Aware Kubernetes Scheduler [I]
40mVirtual Topology for Virtual Machines: Friend or Foe?
22mTransparent MultiNIC Routing for Large AI Models
32mKubernetes Namespaces Unleashed: Unlocking the Full Potential of You...
22mLessons We Learned from Deploying Cloud-Native Telco Cloud for 5G
34mDemystifying Kubernetes Resource Management
04h 00mNx Conf Lite 2022
15mDesigning a Read-Intensive Architecture for Information Systems