1GB Fibre - Hardware 'Good Enough'?


I am switching to VirginMedia (UK) 1Gb fibre connection this week and would like some opinions if existing hardware will be 'good enough', my main consideration is getting stable, fast connection to a hardwired PCs.

I've read the sticky on 500Mb+ connections and the best way is defiantly a Pi4 and my own wifi router. I'll probably end up with that setup.

However, what if I put the supplied Virgin Hub 4 router into modem mode and use my existing Archer C7 as router/wifi (old but reliable). Would that good enough, I dont know if Hub4 has SQM (do I need SQM on both modem and router?)

The Hub 4 is essentially a rebranded Arris’s Touchstone TG3492

Intel Puma 7 (CE2752) CPU, can handle 32Γ—8 bonded channels on DOCSIS 3.0 and 2Γ—2 OFDM and OFDMA on DOCSIS 3.1, supports 3Γ—3 MIMO on 2.4GHz [WiFi]
(802.11n) and 4Γ—4 MIMO on 5GHz (802.11ac Wave 2 – 1.3Gbps)

Virgin are also in the process of rolling out the Hub 5, I'll be getting that in the next couple of months.

Broadband – DOCSIS3.1/3.0 Broadcom 3390S
5 x Antennae
4 x 4 MiMo 5Ghz Broadcom 6715
3 x 3 MiMo 2.4Ghz Broadcom 6710
WiFi 6
1 x 2.5Gbps Broadcom 54991EL
3 x 1Gbps Broadcom 3390S

I would highly recommend going for hardware not connected by using USB and hardware that supports hardware crypto acceleration (ie not the RPI4) especially if "stable" is a requirement but that's your choice.

If the HUB 5 doesn't support "modem mode" it'll be a downgrade and render your planned setup kinda useless. As far as speed goes I highly doubt you'll notice difference between HUB 4 and HUB 5 so there's no need to wait unless it's available for the same price or lower than the current generation.

If you even need SQM (debatable but I'd suggest that you go with it disabled first) you only need to have it running on your router/firewall.

If you're going the SBC route you might also want to consider running another distro/OS that's not aimed for embedded hardware with limited storage as you might want to run software that's not available/easily to package on OpenWrt.

Worth mentioning is that unless you have particular requirements the supplied gateway might actually do the trick on its own.

The Hub 5 has been on trial for some, they are looking to roll out next quarter and its free.

If the Hub 4 is 'good enough' in modem mode, then I dont need to worry about getting a Pi or anything else. However, I thought SQM was 'super important'

Side point - Are we some time off Wifi 6 routers supported properly by Openwrt ? I only see a couple in the list

Only a small (but recently increasing) number of commercial routers offer SQM. Whether you need/want it, is up to you to decide, but if you do the C7 is not going to cut it for your WAN link, sorry. Depending on what you want to achieve with SQM a sngle instance might be enough, but on your router it will have more useful information about packets to allow some pretty nifty isolation modes, while on a modem it will offer much less options. BUT no DOCSIS router offer SQM at all, neither in router or bridged-modem mode, so this is really moot.

Indeed there seems to be a bimodal happiness distribution with USB ethernet dongles, some seem to be pretty reliable and some quite flaky, depends on whether you "feel lucky" whether that is a gamble you are willing to take. But there are other options as well which however increase the price substantially.

About the lack of hardware crypto, I would say, unless you want to use a VPN from your router that can exercise said hardware crypto having or not having it available from the CPUs seems moot. And once we talk VPN one question certainly is will the VPN you want to use actually use the hardware crypto at all.
But these are policy decisions you need to make and hw crypto certainly is not an rpi4B's forte.

One thing the raspberries have going for them is the relative high number of units sold (which means that there can be re-spins if errors are detected like the USB port power issues) but they certainly are not developed with networking use as a priority. Decisions, decisions, decisions...

Here are some plots of a VM 1 Gbps docsis link with and without SQM that give you some idea about what SQM might buy you. Your decision. I do want to add however, that I consider the often heard approach "competent traffic shaping and AQM is not needed above X Mbps link speed" to be not a terribly well founded argument. iMHO you need that if the stock "latency-under-load-increase" of a link is too much for your taste independent of link speed (granted the higher the link speed the harder it is to get hardware that allows competent traffic shaping in software). As an aside, SQM is disabled by default if you install it, since it requires manual setting of some parameters to be useful enough...

In modem mode it will not do anything but be a "dumb" media converter between ethernet and DOCSIS (not exactly true since it just pretends to be dumb, and as so often playing dumb competently is harder than it looks).

That is a decision you need to take/make. I would implore you to at least test whether SQM improves your link's interactive usability under load, but in the end your network, your policy decisions.

I assume by "modem mode" they refer to bridge mode (ie it doesn't do NAT or anything rather acting like a transparent bridge) so if you're going to use that mode you're going to need some kind of router/firewall.

11ax support is at best alpha, you're much better off going for 11ac for now.

I'm starting to wonder if this is some kind of odd behaviour with Linux/OpenWrt in some scenarios or if I'm really lucky. I have only one DOCSIS connection available right now and it's ~150/50 and I don't really see this behaviour at all.

Modem Netgear DG3100 (DOCSIS 3.0 EU) (I know its old but it works just fine)
Testbox: Xunlong Orange Pi PC 2 running FreeBSD 13-STABLE (@ 4x ~620Mhz) - Public IP
Other devices connected: x86 box acting as gateway/firewall (~uses 1mbit or so)

Maxing out the connection increases latency by ~10-15ms (17-18ms idle to 30-32ms tops) that's without any tweaking (newreno as congestion control fwiw) on the testbox.

I'm not really sold on crab (Realtek) NICs, I've been very happy with Intel NICs for many years now although their 2.5G controller did have some issues. I very much prefer the RockPro64 which uses the RK3399and pretty much ticks all boxes. It can possibly also be seen as more open than the Broadcom SoCs that RPi uses as documentation is publically available. That being said, most of experience regarding networknig has been using FreeBSD 13-STABLE which works great (I have one that's been running for 137 days as we speak without any issues) but I do have a few RK3399 devices running Manjaro that also performs well in general.

1 Like

My guess (having experienced bufferbloat on ADSL, VDSL2, and DOCSIS links all with the respective ISPs stock modem-router myself) is that you are simply lucky and/or your ISP competently configured your link (and your link usage falling within the configured profile). But I could be wrong of course.

How do you measure this and how far away are the data sinks/sources?

Yes, I agree that these d not have a terribly good reputation, but the 2 raspberry pi's in use in my network (none running as a router, one ist my GPS-disciplined NTP/PTP time source, the other a raspberry pi 400 serves light desktop duty) are not noticeably hampered by their NICs.

Well, over here it is much harder to source than a raspberry pi 4B... (and I am not sure I think trading two of the A72 cores for 4 A53 is more than neutral) but I agree the specs are pretty attractive otherwise.

Not sure that this is necessarily that important for the crowd that just is looking for a router capable of doing anything interesting at ~1Gbps though.

The story is pretty much the same on any connection that isn't DSL and have a "decent" amount of bandwidth.

I maxed out the connection using curl (http) and tried two remote sites (different ASNs), most likely not the best method but it gives you an idea.

Unless you aren't using USB nics it's actually not Realtek NICs that's being used, do note however that many SBCs uses Realtek PHYs but the controller is usually a variant of DesignWare's networking IP. For occational data transfer they may work ok-ish however.

I'm not sure where you are but Pine64's own shop works really well even in Europe and you have Ameridroid in NA/US so it's usually not that hard to order.

Regarding hardware documentation I guess it doesn't matter too much for the end-user but you're basically relying on the RPi Foundation to fix things which may/may not work in your favour.

Well, I believe you experienced that, but it does not match my own experience. But as I said before, your network, your rules.

What do you mean with "tried two remote sites" and how did you assess "interactivity" during times of high load? Personally I think that flent's RRUL or RRUL_CS8 test is a pretty easy to run tests that allow to assess how latency increases under load in a nice package, but it requires finding some remote netperf servers. But other methods are possible as well, as long as they combine saturating loads (preferably bi-directionally saturating loads) with a more or less continuous latency probing... so in the limit something like running a speedtest against fast.com for 60 seconds (or using curl) while concurrently running sudo mtr -ezb4 -i 0.2 will do the trick....

Well, anything but the 4GB model is out of stock and the pine64 store does not handle duty/VAT, which is a pain as it likely requires a trip to the duty office to get the device and to pay 19% VAT... While I can source rpi4B's from multiple local sellers without getting into duty/VAT issues. I admit though that are just slight inconveniences and not a killer-argument against a rockpro.

Yes, as far as I am concerned however they have a decent track-record so far.... not perfect but good enough.

  • Remote as two other nodes/connections that isn't on the same ASN so in theory latency should be worse due to transit/peering etc.

  • I used ping but ssh worked fine too, not a great metric but you can easily tell when latency increses but I want to stress that I just wanted to see if I could get somewhat similar results as yours without spending a ton of time.

  • I don't see why you'd go for anything less, you're chasing pennies and if you want to use to for something else / more tasks it's very nice to have the ability to do so. The increased amount also gives you more headroom for Rust and Go applications etc =) At least here in worst case you get 25% VAT and a ~8 EUR administration fee which isn't the end of the world and not that far from a RPi 4Gb anyway at least not where I live.

1 Like

Not sure how close you want to get to a gig.
I have 1gig fibre symmetrical and get quite close with a Linksys WRT1900ACSv2

Not quite sure where all the hate has come from for the RPi4 with a USB NIC. From what I've read there's plenty of people using this combo successfully.

The TP-Link UE300 with the RTL8153 chipset seems to be the most popular and reliable option. I'm running a AX88179 USB NIC without issue on a gigabit connection, but this chip seems to give problems for some people.

I've got the DFRobot Router Carrier Board sitting on my desk waiting for a compute module to arrive. Uses an RTL8111 NIC connected via PCIe. Could be an option if you really want to avoid USB.

1 Like

RT3200 does 1Gbit SQM with normal sized MTU, right? Only 80GBP in the UK. So isn't that the winner?

1 Like

I also have the WRT1900ACS and 1Gbps symmetric fiber. I set SQM to:

  • 950000 kbps up/down (actual speed direct from fiber box)
  • Queue discipline: fq_codel
  • Queue script: simplest.qos
  • Link Layer Adaptation = Eth w/overhead, 44 bytes

Now, that's not very hardcore in terms of QoS, and you might think it's not really needed with that amount of bandwidth... But in any case, based on my testing + normal hardcore usage, it works perfectly and balances everything quite nicely at 927 down, 920 up. With any other SQM settings, the speed slows by 100-300 Mbps.

Hope that helps.