Five Ways to Understand Kubernetes Networking for Performance-Intensive Workloads

Explore five ways to implement Container Network Interface (CNI) for performance-intensive Kubernetes workloads in this 36-minute conference talk. Delve into the evolving landscape of Kubernetes networking, focusing on recent innovations that have improved performance for demanding scientific computing and distributed machine learning applications. Learn about different CNI architectures and technologies, their claimed performance advantages, and how to distinguish between them using real-world benchmarks. Gain insights into network considerations for High Performance Computing (HPC) and Artificial Intelligence (AI) workloads, including Remote Direct Memory Access (RDMA), OVN Kubernetes, and Single Root I/O Virtualization (SR-IOV). Examine lab setups, test scenarios, and performance comparisons using tools like Iperf 2, Intel MPI Benchmarks, OpenFOAM for Computational Fluid Dynamics, and genome sequencing basecalling benchmarks.

Intro
High Performance Computing (HPC)
Artificial Intelligence (AI)
AI & HPC - Network Considerations
Remote Direct Memory Access - RDMA
OVN Kubernetes
Single Root I/O Virtualization (SR-IOV)
Legacy OVS / OVN Challenges
Lab Set-Up: K8s Bare Metal
Lab Set-Up: K8s Over OpenStack VM
Test Scenarios
Iperf 2 - TCP Throughput Benchmark
Calico vs Acc OVN: CPU load during test
Accelerated OVN: Flow offloads during test
Intel MPI Benchmarks - PingPong
Computational Fluid Dynamics - OpenFOAM
Genome Sequencing - Basecalling benchmark