[Cake] Re: [PATCH] net/sched: sch_cake: skip redundant cake_mq rate updates

["Jonas Köppeler" <j.koeppeler@tu-berlin.de>]

On 10/25/26 9:38 PM, Shubhang Kaushik (Ampere) wrote:
> cake_mq periodically counts active child qdiscs in cake_dequeue() and
> derives an effective per-child shaper rate. cake_dequeue() currently
> reconfigures every tin after each scan, even when that effective rate
> is unchanged.
> 
> The dequeue-time call chain is:
> 
>    cake_dequeue()
>      -> cake_configure_rates(..., true)
>        -> cake_config_*()
>          -> cake_set_rate()
> 
> With rate_adjust=true, cake_set_rate() only refreshes rate-derived
> state. Tin 0 records the unscaled rate argument in every supported
> CAKE tin mode, so its installed rate identifies whether this call chain
> would change any state.
> 
> Skip the reconfiguration when the effective rate is already installed,
> while continuing to update the activity-sampling timestamp and
> active-queue statistic.
> 
> Fixes: 1bddd758bac2 ("net/sched: sch_cake: share shaper state across sub-instances of cake_mq")
> Signed-off-by: Shubhang Kaushik (Ampere) <sh@gentwo.org>
> ---
> Tested on an 80-core SMP, non-SMT Ampere Altra system (aarch64)
> The test used a 32-TX-queue veth pair, with cake_mq
> bandwidth 2gbit on the transmit side and the peer in a network
> namespace. The workload was:
> 
>    iperf3 -c 198.18.0.2 -P 32 -t 30
> 
> Validation against mainline commit a32e63a613aa80, the patched kernel
> reduced cake_configure_rates() calls by 99.99% (2,477,627 to 206) and
> aggregate helper time by 99.986% (3.477 s to 0.498 ms), while retaining
> approximately 1.91 Gbit/s aggregate throughput (baseline:
> 1.91-1.92 Gbit/s). A separate per-child trace observed no
> same-effective-rate calls on the patched kernel (@unchanged: 0).
> Changing cake_mq bandwidth from 2Gbit to 1Gbit and back invoked
> cake_configure_rates() 64 times, matching all 32 child CAKE qdiscs
> updated for each of the two control-plane changes.
> ---
>   net/sched/sch_cake.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/net/sched/sch_cake.c b/net/sched/sch_cake.c
> index aa4043d662c1cc18858d3f9c80f9d9b281f26f72..dd2539cb472087af9a96b74f2ac571eb8b0a8a27 100644
> --- a/net/sched/sch_cake.c
> +++ b/net/sched/sch_cake.c
> @@ -2052,7 +2052,8 @@ static struct sk_buff *cake_dequeue(struct Qdisc *sch)
>   		if (num_active_qs > 1)
>   			new_rate = div64_u64(new_rate, num_active_qs);
>   
> -		cake_configure_rates(sch, new_rate, true);
> +		if (new_rate != READ_ONCE(q->tins[0].tin_rate_bps))
> +			cake_configure_rates(sch, new_rate, true);
Thanks! Yes, I think the idea is right. I think we could even omit the
READ_ONCE.

Best
Jonas>   		q->last_checked_active = now;
>   		WRITE_ONCE(q->active_queues, num_active_qs);
>   	}
> 
> ---
> base-commit: a32e63a613aa80dfdb9acc0a09efab940e96054f
> change-id: 20261025-net_sched_sch_cake-e6653fd393d6
> 
> Best regards,