NanoPi R4S 4GB Major SQM Issues - OpenWrt 25.12.3

[moeller0]

@adamshumpisxxx

It remained wildly inconsistent as above.

Well, you need to adress the big/little problem first...

[LilRedDog]

@moeller0

No it does not... the problems are manyfold:

Absolutely does, still, provide the information...

SQM is performed on the CPU, as such slower devices may be unable to keep up with your peak internet speed.

Like I make this stuff up.

[phinn]

It seems like people are getting sidetracked and I have doubts there is a regression in 25.12. Have you:

1. Clean install of 25.12.5. Apply packet steering and/or SFO.
2. Applied StarWhiz's settings in case it is not included as he thought. He gets 780Mbps with that. This sort of tweak is always needed with hybrid cores unfortunately.
3. Tried SQM with cake_mq. There is still no NanoPi / Rockchip users that replied on that thread with results so you could be the first.

[moeller0]

@LilRedDog

Absolutely does, still, provide the information...

Could you please elaborate, what does? and what/who provides that information for required CPU cycles?

@LilRedDog

Like I make this stuff up.

Puzzled, how is that incompatible with it being hard to reliably estimate a number of required CPU cycles for X Mbps traffic shaping? No objection that SQM is a bit of a CPU hog, but that does not mean that I can tell things like required CPU frequency...

I feel we are talking past each other, resulting in avoidable friction...

[adamshumpisxxx]

My service IS NOT symmetrical and as above in the OP asymmetrical 920-940 / 40 gigabit service. Same provider. Same equipment (MB8611 modem). Same "service". Different tier. Line tech already came and mitigated any and all dBmV / SNR issues I was having on the downstream after regional "upgrades" to the service. Modem is now in perfect working parameters and runs like a Swiss watch. I don't want to make this sound like a new issue worth considering as a potential problem because everything was settled back in April. The tier upgrade, however is brand new.

I am confident this is not a service issue. I am getting the same amount of latency and bufferbloat without SQM as per Waveform now as I was pre-upgrade of which I was able to fully mitigate previously with what I consider to be PLENTY of headroom as per what the htop / netdata metrics showed.

[adamshumpisxxx]

Thank you for this. I believe I will start here. Questions first...

What exactly would you recommend for both packet steering and SFO?

Would StarWhiz's tweaks not be redundant / would they not mess something up if applied on top of the firmware which is claimed to have already included them?

Actually...to end the speculation...can I get some confirmation from someone somewhere that said optimizations have actually been made? I'm not convinced this was ever a thing at this point and I'm not sure how to prove it one way or another. It may have also been removed in more recent updates if such tweaks in the firmware were ever made.

As for cake_mq this is the first I'm hearing about it so if it is an option I will give it a shot. Thank you!

[LilRedDog]

I see you did not open the hyperlink the Wiki added.

It is so detailed, leaving it in was probably seen as untidy.

-edit

@moeller0

I feel we are talking past each other, resulting in avoidable friction...

Not only did you mistakenly dismiss 'such an idea', you doubled down, yes?

[moeller0]

@LilRedDog

Not only did you mistakenly dismiss 'such an idea', you doubled down, yes?

Where in the forum is a note that traffic shaping X Mbps takes Y cpu cycles? Really puzzled... I am not doubting that traffic shaping is CPU intensive, but that was not your question was it? You wanted to know "the number of cycles, 1G needs?" which neither the wiki nor the forum definitively tell... so yes I am doubling and tripling down that the answer you seek is not available, after explaining a few reasons why expecting something like this to be quite optimistic. (But note the detailed SQM wiki has sections explaining how to monitor CPU load).

[moeller0]

@adamshumpisxxx

What exactly would you recommend for both packet steering and SFO?

Packet steering should be happening only on the two beefy CPUs.
SFO should be disabled at first, as depending on the way the WAN interface is used it might or might not work with SQM, so I would only try to enable that after getting the basic sqm config working.

[LilRedDog]

We are talking past each other:

I was replying to a 'maybe it is not' post and I was pointing out it has been quantified, and is in the Wiki.

[LilRedDog]

@moeller0

so yes I am doubling and tripling down that the answer you seek is not available,

@dlakelan

It's such a frequently asked question over the last few years that we really need a post we can point people to.

So your cable company rolled out 500Mbps or 1Gbps download speeds, or you have 1Gbps symmetric fiber (GPON/EPON) from ATT or Centurylink or Orange or whoever. You realize your old all in one router from 2009 is not up to the task of handling this, but hey by now stuff should be cheap and available that will handle your new fiber connection right? So you want advice about a router for between $50 and $100 that will route your symmetric gigabit connection while maintaining low latency using SQM and also easy to flash, and you're used to all-in-one devices with wifi, so it should do it all... can someone offer you a selection of 4 or 5 of them to check out for availability in your area?

The answer is NO. Why? Because they don't really exist.

Let's take a look at the math: At 1Gbps using 1500 byte packets, you need to send/receive 83333 packets per second. The packets need to be received by an interrupt, go through the firewall, be inspected, maybe have NAT applied, sent into a queue, the queue calculates rates to avoid over-sending on the link and causing buffers, and then hardware interrupts are serviced to actually send the packet along...

At 1 GHz processing rate, each packet gets 12000 clock cycles of calculation if the CPU is maxed out doing nothing but processing packets.

Evidently in an ideal world, we should have maybe 1.2GHz processors or better, and maybe have two cores at least one can handle interrupts on the receive interface, and one can handle interrupts on the send interface, and they can share the firewall and queueing duties. Let's not forget that there's RAM latency and bandwidth issues if the packets need to go from kernel to userland (like for OpenVPN) and encryption/decryption also for VPNs.

An obvious comparison point is something like the ZBOX Edge CI341 mini PC:
N4100 processor from 1.1 to 2.4GHz with dual LAN ports. It lists on Amazon for $179 without ram, SSD, or Operating system. You can probably turn this into a router for an additional $60 in RAM and SSD and installing OpenWrt, so your final price is $240 for a wired only router that will be quite competent. Now all you need to do is buy $40-120 worth of smart managed switch, and two TP-Link EAP access points, or ubiquiti access points, or maybe some older all-in-one routers running OpenWrt in Dumb AP mode. Let's budget $60 for each access point, and your house needs two of them... so we have:

• x86 Router: $240
• 8 port Smart Switch: $40
• two APs: $120

Total: $400

This is the order of magnitude you should expect to spend to get very good performance on your new Gigabit symmetric line.

Can you do it cheaper? Yes. For example the Raspberry Pi 4 is very popular now. It's more of a development board than an all in one package. The real world performance numbers show that it can route and SQM gigabits of packets using 25% of its CPU capacity or so. If you buy the 2GB version and a case and power supply and the UE 300 USB ethernet dongle, it winds up being closer to $120 than the $240 for the above x86 computer. It's widely available and a good choice, but you're still in the range of $250 after buying your switch and a couple access points. This is more or less the budget version of a good setup. There are also some other good candidate boards which are less widely available but probably very good candidates. the NanoPi R4S or the RockPro64 come to mind.

But at the end of the day, as you move above 500Mbps you should consider the idea that you now have a serious bit of computing to do just to route and firewall and SQM your network, and you'll be better off with a component based network rather than sticking to your all-in-one "wireless router", at least maybe until someone comes out with a good line of multi-gigahertz multi-core multi radio all-in-one devices for 1/4 the cost of a component wise setup. Don't hold your breath over the next 2 years.

In the link.

What is going on?
Are you saying the linked write up is wrong?
Just pulled out of a hat?

[moeller0]

@LilRedDog

We are talking past each other:

Ah, I guess we agree I will stop now.

[adamshumpisxxx]

Updated to 25.12.5, set Packet Steering to Enabled All Cores, enabled Software Flow Offloading, set the interface to eth0, set ingress to 840000, set egress to 39000, left the default Cake / Piece of Cake for Queue Discipline, set the Link Layer to Ethernet with overhead, set the Per Packet Overhead to 42, enabled SQM and hit apply.

This is the best I'm going to get so I'm inclined to leave it that way. I have NO IDEA how StarWhiz with an asymmetrical gigabit connection got as good as he got (800+ Mbps SQM on 1200 / 40 service as per his claims on his GitHub) and at this point I no longer care. I'm no longer inclined to believe anyone claiming anything regarding this device without screenshots and proof of ownership / active use after all of this. Thanks to those who tried to help and those who actually did.