Memory Recycling for Network Interfaces
summary
Proposes a page_pool-based memory recycling API that gives drivers DMA mappings automatically and unifies XDP and skb paths, implemented as skb free interception now and being moved to moves recycling info into struct page for fragment support. Tests on a Cortex‑A53 Marvell board show >2x throughput on specific small‑packet workloads with a big performance step at 257 bytes, measured on non‑IOMMU hardware. Tradeoffs: higher reserved RAM, pool sizing risk, and pending MLX5 validation. supports XDP and skb paths
AI-written summary · about
chapters (10)
- 0:00Introduction to memory recycling
- 1:14Motivation for new API
- 3:19Current network buffer lifecycle
- 5:03Proposed recycling mechanism
- 6:51Challenges with current approach
- 8:13Performance test results
- 11:49Limitations of page pool
- 13:15XDP compatibility benefits
- 15:30Future testing and validation
- 18:45Next steps and Q&A
related talks
33mZero Copy Receive using io_uring
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
16mQueue-Mem: Energy-Efficient Hardware Storage for Advanced Network Function Acceleration
26mRunning DPDK on Non-Cache Coherent Platforms: Challenges and Solutions
43mLessons Learnt from Reusing QDMA NIC to Base Band PMD
11mSupporting Hypervisor Kexec with Modern Devices
18mSTYX: Exploiting SmartNIC Capability to Reduce Datacenter Memory Tax
44mLinux Networking with MANA: RX Path Optimization and Netsh...