https://peer.madiator.cloud/videos/watch/2a4678de-d47e-49a2-8bdd-15e8caf137d5 Recording of Qcon 2026 talk: "Accelerating Container Networking: Fun with Fastpathing". Abstract and slides from this talk available at https://qconlondon.com/presentation/mar2026/accelerating-container-networking-fun-fastpathing In this talk from QCon, Runpod Cloud Engineer Murat Magomedov explores how modern container networking can take advantage of the high-speed networking hardware driving today's AI workloads. He introduces RDMA (Remote Direct Memory Access) and RoCE (RDMA over Converged Ethernet), explaining how these technologies bypass the kernel networking stack to achieve line-rate throughput on 400 Gbps NICs—essential for distributed model training and multi-node inference where TCP simply can't keep up with the volume of weight transfers required for collective reduction operations. The talk walks through the practical challenges of enabling RDMA inside containers, starting with why traditional Docker bridge networks and veth pairs break kernel-bypass traffic. Murat demonstrates progressively more sophisticated approaches: host networking (not really containerized), MACVLAN/IPVLAN drivers (fast but with no isolation), and VXLAN overlays (isolated but initially blocked by the bridge). The key insight is that by stacking Linux virtual interfaces—for example, attaching a MACVLAN to a VXLAN endpoint rather than directly to the physical NIC—you can preserve isolation while still getting the NIC to accept a hardware offload, verified by inspecting the GID table with tools like show_gids and ibv_devinfo. The final sections cover multi-NIC setups common in modern GPU servers (typically one NIC per GPU in rail-optimized leaf-spine topologies) and the surprising result that even bonded interfaces can work with RDMA, with the NIC accepting many-to-one hardware offloads to the same logical device. Murat closes by noting that most Kubernetes CNIs likely leave significant performance on the table, since enabling these offloads often requires only adding a dev argument to standard ip command and he encourages viewers to experiment with stacking Linux drivers in ways the documentation never quite spells out. 0:00 - Introduction 0:33 - Why RDMA matters for AI workloads 1:51 - Kernel stack vs. RDMA user-space path 2:51 - Infiniband, RoCE v1/v2, and Ultra Ethernet history 4:42 - Inspecting RDMA devices and GIDs 7:03 - Baseline performance test with ib_write_bw 8:19 - Container networking: bridges and veth pairs 9:34 - Multi-node L2 setups and their limits 11:58 - Host networking (the non-solution) 12:38 - MACVLAN and IPVLAN drivers 14:47 - Offloading a MACVLAN to the NIC 16:28 - MACVLAN shortcomings: no isolation 16:54 - VXLAN overlays and hardware offload 18:38 - Why the bridge still blocks RDMA 20:30 - Stacking MACVLAN on top of VXLAN 23:57 - Verifying the offload chain works at line rate 25:44 - Multi-NIC setups and rail-optimized topologies 27:25 - Bonding NICs for RDMA (yes, it works) 31:23 - Conclusions and the Linux stacking philosophy 33:19 - Wrap-up Tue, 21 Jul 2026 20:54:33 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://peer.madiator.cloud https://peer.madiator.cloud/lazy-static/avatars/e2927aea-9e0a-454c-9327-68ec87b4da11.png https://peer.madiator.cloud/videos/watch/2a4678de-d47e-49a2-8bdd-15e8caf137d5 All rights reserved, unless otherwise specified in the terms specified at https://peer.madiator.cloud/about and potential licenses granted by each content's rightholder.