Re: [PATCH net-next v7 00/11] net: tunnel: introduce noref xmit flows for tunnels

[Marek Mietus <mmietus97@yahoo.com>]

W dniu 3.02.2026 o 09:32, Paolo Abeni pisze:
> On 1/27/26 8:04 AM, Marek Mietus wrote:
>> Currently, tunnel xmit flows always take a reference on the dst_entry
>> for each xmitted packet. These atomic operations are redundant in some
>> flows.
>>
>> This patchset introduces the infrastructure required for converting
>> the tunnel xmit flows to noref, and converts them where possible.
>>
>> These changes improve tunnel performance, since less atomic operations
>> are used.
>>
>> There are already noref optimizations in both ipv4 and ip6.
>> (See __ip_queue_xmit, inet6_csk_xmit)
>> This patchset implements similar optimizations in ip and udp tunnels.
>>
>> Benchmarks:
>> I used a vxlan tunnel over a pair of veth peers and measured the average
>> throughput over multiple samples.
>>
>> I ran 100 samples on a clean build, and another 100 on a patched
>> build. Each sample ran for 120 seconds. These were my results:
>>
>> clean:      71.95 mb/sec, stddev = 1.71
>> patched:    74.92 mb/sec, stddev = 1.35
> 
> Which H/W are you using? I've never seen so low figures in this decade,
> possibly not even in the previous one. Expected tput on not too
> obsoleted H/W is orders of magnitude higher, and we need more relevant
> figures.
> 

Sorry for the late response.

I tested this on my i9-13900. The reason for these low figures is that I
used "iperf3 -l 2 -c <IP>" to measure throughput. I manually lowered the
packet length because I figured there might be some other factors at play
that could add noise to my measurements. I'll measure again without this
limit. From initial testing, it looks like this performance gain is also
achieved with the default packet size.

I'll include new measurements along with v8.

> Thanks,
> 
> Paolo
> 
>