On Sat, 31 Jan 2026 17:40:15 +0800 Mike Chang <[email protected]> wrote:
> Hi DPDK users, > > I’d appreciate feedback on SPD’s Greedy Reshaper control strategy: > > Context > - Input each tick: per-worker RX KPPS from a perf core; small move budget > per interval. > - Action: scan software RETA (256) and flip entries hot→cold, bounded by > max_moves, to reduce imbalance. > - Aim: improve fairness/stability without disruptive global remaps. > > What it is > • Software-only & portable: no NIC-specific features; all reshaping is done > in user space. Suit for SDN. > • Bounded, in-place edits: each interval flips a small number of RETA > entries to move hot buckets from overloaded to cold workers, keeping > overhead predictable. > > Questions > 1) Heuristics: any pitfalls with bucket flipping under bursty traffic? > Better ranking inputs you’d suggest? > 2) Timing: practical sampling intervals on your x86/Arm setups to keep > overhead hidden? > 3) Validation: happy to compare against your traces; SPD logs per-second > CSV. > > Repo/Docs: https://github.com/mikechang-engr/software-packet-distributor > Testbed so far: LX2160A-RDB, LSDK 21.08, Linux 5.10.35, DPDK 19.11.7 > (PCAP/NULL vdev). > > Thanks, > Mike More cores does not mean more performance. Every time you move packet between cores you incur another cache miss. The code style is unreadable for Linux users; no indentation or line breaks. Is this a candidate for oobfuscated or vibe coding? Get on a recent version of DPDK, not one that is 6 years old.
