Date: Mon, 19 Feb 2024 20:01:58 +0100
From: Toke Høiland-Jørgensen <toke@...hat.com>
To: Sebastian Andrzej Siewior <bigeasy@...utronix.de>
Cc: Jesper Dangaard Brouer <hawk@...nel.org>, bpf@...r.kernel.org,
 netdev@...r.kernel.org, Björn Töpel
 <bjorn@...nel.org>, "David S. Miller"
 <davem@...emloft.net>, Alexei Starovoitov <ast@...nel.org>, Andrii
 Nakryiko <andrii@...nel.org>, Daniel Borkmann <daniel@...earbox.net>, Eric
 Dumazet <edumazet@...gle.com>, Hao Luo <haoluo@...gle.com>, Jakub Kicinski
 <kuba@...nel.org>, Jiri Olsa <jolsa@...nel.org>, John Fastabend
 <john.fastabend@...il.com>, Jonathan Lemon <jonathan.lemon@...il.com>, KP
 Singh <kpsingh@...nel.org>, Maciej Fijalkowski
 <maciej.fijalkowski@...el.com>, Magnus Karlsson
 <magnus.karlsson@...el.com>, Martin KaFai Lau <martin.lau@...ux.dev>,
 Paolo Abeni <pabeni@...hat.com>, Peter Zijlstra <peterz@...radead.org>,
 Song Liu <song@...nel.org>, Stanislav Fomichev <sdf@...gle.com>, Thomas
 Gleixner <tglx@...utronix.de>, Yonghong Song <yonghong.song@...ux.dev>
Subject: Re: [PATCH RFC net-next 1/2] net: Reference bpf_redirect_info via
 task_struct on PREEMPT_RT.

Sebastian Andrzej Siewior <bigeasy@...utronix.de> writes:

> On 2024-02-15 21:23:23 [+0100], Toke Høiland-Jørgensen wrote:
>> The tricky part is that the traffic actually has to stress the CPU,
>> which means that the offered load has to be higher than what the CPU can
>> handle. Which generally means running on high-speed NICs with small
>> packets: a modern server CPU has no problem keeping up with a 10G link
>> even at 64-byte packet size, so a 100G NIC is needed, or the test needs
>> to be run on a low-powered machine.
>
> I have 10G box. I can tell cpufreq to go down to 1.1Ghz and I could
> reduce the queues to one and hope that it is slow enough.

Yeah, that may work. As long as the baseline performance is below the
~14Mpps that's 10G line rate for small packets.

>> As a traffic generator, the xdp-trafficgen utility also in xdp-tools can
>> be used, or the in-kernel pktgen, or something like T-rex or Moongen.
>> Generally serving UDP traffic with 64-byte packets on a single port
>> is enough to make sure the traffic is serviced by a single CPU, although
>> some configuration may be needed to steer IRQs as well.
>
> I played with xdp-trafficgen:
> | # xdp-trafficgen udp eno2  -v
> | Current rlimit is infinity or 0. Not raising
> | Kernel supports 5-arg xdp_cpumap_kthread tracepoint
> | Error in ethtool ioctl: Operation not supported
> | Got -95 queues for ifname lo
> | Kernel supports 5-arg xdp_cpumap_kthread tracepoint
> | Got 94 queues for ifname eno2
> | Transmitting on eno2 (ifindex 3)
> | lo->eno2                        0 err/s                 0 xmit/s
> | lo->eno2                        0 err/s                 0 xmit/s
> | lo->eno2                        0 err/s                 0 xmit/s
>
> I even tried set the MAC address with -M/ -m but nothing happens. I see
> and on drop side something happening when I issue a ping command.
> Does something ring a bell? Otherwise I try the pktgen. This is a Debian
> kernel (just to ensure I didn't break something or forgot a config
> switch).

Hmm, how old a kernel? And on what hardware? xdp-trafficgen requires a
relatively new kernel, and the driver needs to support XDP_REDIRECT. It
may be simpler to use pktgen, and at 10G rates that shouldn't become a
bottleneck either. The pktgen_sample03_burst_single_flow.sh script in
samples/pktgen in the kernel source tree is fine for this usage.

-Toke

Powered by blists - more mailing lists