The case for zero-copy in containers: Accelerating...
summary
Built a kernel-side queue leasing mechanism that lets a container bind virtual NIC queues which proxy to a physical NIC queue, enabling AF_XDP and memory-provider zero-copy (io_uring / devmem) from inside Pods. Implementation adds a new q_create/lease API, lease tracking and netdev lock ordering, and a prototype used in KubeVirt showing about 37% throughput improvement and 30–35% lower latency; current work targets RX queues first, a TX hook, and a proposed BPF multi-attach / queue-range XDP API.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 6:05Memory providers in containers
- 11:05Zero-copy queue leasing proposal
- 27:36XDP TX hook and BPF multilink
- 38:41Q&A session
- 44:11Technical clarifications and closing
related talks
51mAccelerate containers and VMs networking with vDPA
33mZero Copy Receive using io_uring
01h 21mTutorial: Developing High-performance Networking Applications
26mFreeFlow: Software-based Virtual RDMA Networking for Containerized Clouds
46mNetwork Traffic Quality of Service (QoS) Classes for Containers & Pods Via EBPF/XDP
34mTroubleshooting Low Latency Application on CNF Deployment
14mHigh-throughput and Flexible Host Networking for Accelerated Computing
28mLow-Latency, Deterministic Networking with Standard Linux using XDP Sockets