[Cerowrt-devel] Fwd: Realtek RTL8156 devices defaulting to CDC-NCM instead of vendor mode, resulting in reduced performance

[Dave Taht]

.6ms considered good.

---------- Forwarded message ---------
From: Forest Crossman <cyrozap at gmail.com>
Date: Mon, May 2, 2022 at 3:49 PM
Subject: Realtek RTL8156 devices defaulting to CDC-NCM instead of
vendor mode, resulting in reduced performance
To: <hayeswang at realtek.com>, <davem at davemloft.net>, <kuba at kernel.org>
Cc: <netdev at vger.kernel.org>, <linux-usb at vger.kernel.org>


Hi, all,

I recently purchased a pair of USB to 2.5G Ethernet dongles based on
the RTL8156, and have so far been very happy with them, but only after
adding some udev rules[0] to to take advantage of the r8152 driver by
switching the devices from their default CDC-NCM mode to the vendor
mode. I was prompted to use those rules to switch the driver because
one of the adapters (based on the RTL8156A) would get very hot, up to
120 F (49 C) even while idle, and the round-trip latency directly
between the pair of adapters was about 3 ms, and I couldn't help but
wonder if maybe the vendor mode might be more efficient.

After performing some tests of latency and power consumption, testing
first with both adapters in NCM mode and then again with both in
vendor mode, I proved my hunch correct. I discovered that, in a
disconnected state, the RTL8156A adapter used about half as much power
(0.64 W -> 0.30 W) while the RTL8156B adapter saw a 21% reduction in
power (0.34 W -> 0.27 W). Similarly, in a connected-but-idle state the
RTL8156A again saw about a 55% savings in power consumption (2.17 W ->
0.97 W) and a 40% savings in the RTL8156B adapter (0.94 W -> 0.56 W).
It was only under full load that the fewest power savings were seen,
with a reduction of only 15% in the RTL8156A (2.23 W -> 1.90 W) and no
savings for the RTL8156B (0.96 W). Similarly, round-trip latency while
idle went from 3 ms to 0.6 ms. I also tested under load and saw much
larger latency savings and reduced packet loss, but forgot to write
down the numbers (I can run the tests again if someone really wants me
too). Also, jumbo frames drastically reduced performance under NCM
mode, while vendor mode handled it like a champ (again, I forgot to
write down the numbers but can test again if asked).

So, with all the benefits I've seen from using these adapters in their
vendor mode, is there still a reason to let the kernel prefer their
NCM mode? It'd be nice to be able to get the maximum performance from
these adapters on any Linux system I plug them into, without having to
install a udev rule on every one of those systems.

If anyone would like to try replicating the results I listed here, or
to perform new tests, the specific RTL8156A adapter I used is the
Ugreen CM275[1] and the RTL8156B adapter is the Inateck ET1001[2].


Curious to hear your thoughts on this,

Forest


[0]: https://github.com/bb-qq/r8152/blob/160fb96d2319cdf64ae7597e8739972934ac83b2/50-usb-realtek-net.rules
[1]: https://www.amazon.com/gp/product/B081TY1WQX/
[2]: https://www.amazon.com/gp/product/B08VN3DGK6/


-- 
FQ World Domination pending: https://blog.cerowrt.org/post/state_of_fq_codel/
Dave Täht CEO, TekLibre, LLC