Re: genet0 speeds under NetBSD/aarch64 on Rpi4 8 GB using UEFI/ACPI v1.17 firmware

[Paul Mather <bsd-lists%gromit.dlib.vt.edu@localhost>]

On Jul 23, 2020, at 4:47 PM, Jason Mitchell <jmitchel%bigjar.com@localhost> wrote:

> On 7/23/20 11:53 AM, Paul Mather wrote:
>> 
>> After my original message I posted an addendum (https://mail-index.netbsd.org/port-arm/2020/07/23/msg006887.html) that was an iperf3 session run against a different iperf3 server (my pfSense gateway).  That one got similar speeds to yours above.
>> 
>> It's not clear to me why receiving from a quiescent laptop was a lot slower than from my pfSense gateway, but there you go.  (Note: receiving speed was not slow on the Raspberry Pi OS system.)

> I don't know enough to comment authoritatively, but through trial and error I've found that certain chipsets and adapters perform better than others when using iPerf. An good example of poor performing chips are those from Realtek. They're cheap, but it seems like they require a lot of processing power. Meaning if you have loads of CPU, they're fine, but otherwise they are slower than a better chipset (Intel for example). I'm not certain about the why, this is just what I've found.


I agree that some NICs are better than others, and the same goes for driver implementations.  In this case, the differences were at the NetBSD end, not the laptop end.  By that I mean the laptop was not the limiting factor: it was able to achieve GbE speeds sending/receiving with the RPi4 when it was running Raspberry Pi OS.  It was when the RPi4 was running NetBSD that the speeds dropped.


> Also, OS can make a difference. Windows is quite bad at iPerf performance, unsurprisingly. I'd be curious to know what kind of chipset the laptop is using and how that chipset is attached to the CPU (e.g. PCIe, USB, etc.)


The laptop is an old Apple laptop.  The Ethernet controller is on the PCI bus.  I looked up the device IDs (0x10de/0x0ab0) and it said it is an NVIDIA Corporation MCP79 Ethernet device.


> Finally, adding -w 227k can cause a big difference in performance. Testing right now I've seen everything from .3% (342Mbit vs 341Mbit) to roughly a 33% increase (410Mbit vs 310Mbit) when adding -w 227k on the client side. That's a large a window as you can set on NetBSD, FreeBSD, and Linux (maybe 227k is the largest the application supports, I don't know).


Thanks for the tips.  I was hoping I would be able to achieve GbE max speeds without any special tuning.  That is what seemed to happen under Raspberry Pi OS.  As someone observed elsewhere in this thread, it's likely the driver in Raspberry Pi OS is implementing hardware offloading and other things that improve throughput.

Cheers,

Paul.