Linux Networking with MANA: RX Path Optimization and Netsh...
summary
MANA's RX path changes allocation from one page per packet to page_pool fragments plus pre-DMA-mapped page pools per RX queue, letting up to 35 1500‑byte buffers share a 64 KB page and cutting per-packet overhead by ~97% (per-Q memory drops from ~512 MB to ~14.7 MB; total RX footprint from ~16.4 GB to ~470 MB), boosting single‑TCP throughput up to 200% throughput on 64 KB ARM64 (demo: ~20 Gbps → ~60 Gbps) and ~15% on x86_64; driver integration uncovers and fixes deadlocks and missing netdev_ops, and driver-level shaping uses netshaper YAML Netlink with new iproute2 commands.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:04What is MANA
- 2:30RX path problem
- 7:26RX path details
- 11:21RX path solution
- 13:09Performance impact
- 18:06Demo RX path
- 21:19Bandwidth control intro
- 21:41Problem statement
- 22:45Previous approaches
- 24:57Net shapers overview
- 26:23Net shapers commands
- 27:09Demo net shapers
- 31:46Challenges and fixes
- 33:31Future plans
- 35:39Q&A
related talks
43mLessons Learnt from Reusing QDMA NIC to Base Band PMD
23mMulti-Path Transport for RDMA in Datacenters
20mPostMan: Rapidly Mitigating Bursty Traffic by Offloading Packet Processing
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking
17mRearchitecting the TCP Stack for I/O-Offloaded Content Delivery
34mCan the Linux networking stack be used with very high speed applications?
02h 53mPreview Sessions
20mNetKernel: Making Network Stack Part of the Virtualized Infrastructure