RE: Slow 'real world' network performance

[Marcin Gondek <drixter%e-utp.net@localhost>]

Hi

Sorry but Outlook force to top-posting :-/

I've checked also how this looks on my system and result wasn't ok, similar to your ones, ~200-300Mbit/s. I didn't look on that last time and also my NIC and CPU was different.
Looker deeper once I've replaced my NIC with 10Gbit I've little scanned what I can do, with below ones I'm able to close 900Mbit of my Internet connection.

I have no place to simulate 10Gbit full E2E, I have only one server with 10Gbit.

devil# wget -4 --no-proxy https://waw-pl-ping.vultr.com/vultr.com.1000MB.bin
--2025-10-25 19:43:04--  https://waw-pl-ping.vultr.com/vultr.com.1000MB.bin
Resolving waw-pl-ping.vultr.com (waw-pl-ping.vultr.com)... 70.34.242.24
Connecting to waw-pl-ping.vultr.com (waw-pl-ping.vultr.com)|70.34.242.24|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 1048576000 (1000M) [application/octet-stream]
Saving to: 'vultr.com.1000MB.bin'

vultr.com.1000MB.bin                                        100%[========================================================================================================================================>]   1000M  99.5MB/s    in 10s

2025-10-25 19:43:14 (98.0 MB/s) - 'vultr.com.1000MB.bin' saved [1048576000/1048576000]

devil#

What was added, strictly related to network interface.

  # Interface-specific tuning for ixg0
    hw.ixg0.rx_process_limit=512
    hw.ixg0.tx_process_limit=512
    hw.ixg0.enable_aim=1
    hw.ixg0.num_tx_desc=4096
    hw.ixg0.num_rx_desc=4096

    # Interrupt moderation per queue (adjust based on workload)
    hw.ixg0.q0.interrupt_rate=75000
    hw.ixg0.q1.interrupt_rate=75000
    hw.ixg0.q2.interrupt_rate=75000
    hw.ixg0.q3.interrupt_rate=75000
    hw.ixg0.q4.interrupt_rate=75000
    hw.ixg0.q5.interrupt_rate=75000
    hw.ixg0.q6.interrupt_rate=75000
    hw.ixg0.q7.interrupt_rate=75000
    hw.ixg0.q8.interrupt_rate=75000
    hw.ixg0.q9.interrupt_rate=75000
    hw.ixg0.q10.interrupt_rate=75000
    hw.ixg0.q11.interrupt_rate=75000

Thanks,

--
Marcin Gondek / Drixter
http://fido.e-utp.net/
AS56662

-----Original Message-----
From: netbsd-users-owner%NetBSD.org@localhost <netbsd-users-owner%NetBSD.org@localhost> On Behalf Of Peter Miller
Sent: Saturday, October 25, 2025 7:29 PM
To: Michael van Elst <mlelstv%serpens.de@localhost>
Cc: netbsd-users%netbsd.org@localhost
Subject: Re: Slow 'real world' network performance

On Sat, Oct 25, 2025 at 2:19 AM Michael van Elst <mlelstv%serpens.de@localhost> wrote:
> The NetBSD defaults are from another time. Tuning the settings is the 
> first requirement on modern networks.
>
> But your "real world" is probably the internet. There performance 
> depends a lot on congestion control and error recovery, and the NetBSD 
> code is old and "conservative".

I have never tuned my oses before. How does one go about finding where the bottlenecks are? Are there any good resources you can recommend? I haven't started investigating this yet, but will look later when I have some time.

I should point out the obvious info I forgot. This particular test server is running NetBSD 10.1
4 cores of an Intel E5-2680 v4
8 gigs of ram
It has a 10 Gbit shared port, using vioif0 I'm not sure what hard drive is, but it's 50 Gig and shows up as Qemu.
It's also no more than 250 miles from where I live and am testing from.

(Apologies for top posting earlier)
--
Thanks
Peter