On 26 Jan 2026 at 07:45:29 PM, Théo Lebrun [off-list ref] wrote:

On Mon Jan 26, 2026 at 3:29 PM CET, Andrew Lunn wrote:

quoted

quoted

quoted

I was more interested in plain networking, not XDP. Does it perform
better with page pool? You at least need to show it is not worse, you
need to avoid performance regressions.

I retested with iperf3. The target has a single rx queue with iperf3
running with no cpu affinity set.

|              |  64 | 128 |
| baseline     | 273 | 545 |
| pp (page)    | 273 | 544 |
| pp (2 frags) | 272 | 544 |

So no real difference. That is unusual, it is typically faster, or if
it is always doing line rate, it uses less CPU time. That might
suggest the page pool integration is not optimal?

One more data point. I get line rate with & without page_pool so below
are CPU times from /proc/stat:

         upstream    pp
user            1     1
system        179    91 (!!!)
idle         7874  7303
softirq        35    37

16K pages on Mobileye EyeQ5 (MIPS), 7 fragments per page.

Paolo shared 64 versus 128 measurements but I am unsure what those stand
for; I doubt it can be packet size as xdp-bench does not have it as a
parameter. https://man.archlinux.org/man/extra/xdp-tools/xdp-bench.8.en

64 and 128 are packet size in bytes.
For the first test I used xdp-trafficgen on the sender side and
xdp-bench (skb-mode) to count the drops in pps on my board.

For the stack test I used iperf3 (UDP) similarly with 64 and 128 for the
length option.

Measurement incantation:

   cat /proc/stat > /tmp/a && \
   iperf3 -c $IP && \
   cat /proc/stat > /tmp/b && \
   awk 'NR==FNR && $1=="cpu" {user=$2;sys=$4;idle=$5;softirq=$8;next}
        $1=="cpu" {printf "user\t%5d\n", $2-user}
        $1=="cpu" {printf "system\t%5d\n", $4-sys}
        $1=="cpu" {printf "idle\t%5d\n", $5-idle}
        $1=="cpu" {printf "softirq\t%5d\n", $8-softirq}
   ' /tmp/a /tmp/b

Thanks,

--
Théo Lebrun, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com