GRO Library Enhancements
summary
Claims a hash-based 5-tuple lookup replaces iterative table scans and cuts RT gr_reassemble CPU from ~30% to ~18% at 32 flows and from ~33% to ~15% at 64 flows while single-flow cost stays similar. Demonstrates changes in a testpmd VM testbed with client/server traffic, BPF MAC/checksum rewriting and throughput measurements. Proposes per-tuple configurable timeout via an rte_flow-like masked hash table so applications program timeouts, and also adds fixed push-flag reordering behavior and IPv6 support.
AI-written summary · about
chapters (6)
- 0:00Introduction and GRO basics
- 1:27Current GRO implementation issues
- 3:58Hash-based flow search validation
- 6:31Per-tuple timeout proposal
- 9:29Packet reordering fix and IPv6 support
- 12:02Q&A on flow detection
related talks
25mDebug Functional and Performance Issues in Rte_flow
17mPyrrha: Congestion-Root-Based Flow Control to Eliminate Head-of-Line Blocking in...
25mApproximating Fair Queueing on Reconfigurable Switches
28mSafe sharing of the network with eBPF
31mFlow-based Packet Processing Framework on DPDK and VPP
14mMultitenancy for Fast and Programmable Networks in the Cloud
16mPacket Order Matters! Improving Application Performance by Deliberately Delaying Packets
23mACK to the Future: TCP in 2025