Which router for 1000Mbit WAN

changed cpu clock 720->1000Mhz. everything else is as it comes with snapshots (+flow offload enabled)

edit: just figured this answers @jeff 's question: since the increased clock reflects in higher throughput, all flow offloaded packets must be going through cpu so the SQM and/or packet inspection should be possible

Disclaimer for all the people who only read this and don't discuss here: don't do this except in pure sandbox play environments if you want stability and hardware longevity.

39% overclock on a generation 2014 single core MIPS device will not give this aging device the performance of today's routers, but for sure it will risk stability of your network service and longevity of the overclocked device if you don't do hardware modifications and you know exactly what you do.

Also many tasks dont scale 1:1 with CPU clock speed. Assume the task only benefits 50% of more clock speed then you only get 19% more performance. This is often not perceptible for real time tasks. But stability problems are perceptible if they occur.

Another from my point of view more sane option for more router performance would instead be buying a recent multi core ARM architecture new router. There are plenty of models with OpenWrt support, which are also available even cheaper on the used market, e.g.:
ipq40xx: https://downloads.openwrt.org/snapshots/targets/ipq40xx/generic/
ipq806x: https://downloads.openwrt.org/snapshots/targets/ipq806x/generic/
mvebu: https://downloads.openwrt.org/snapshots/targets/mvebu/

3 Likes

IMHO this is a policy decision individual network operators need to make, it does not help to either a priori declare SQM to be required or unnecessary. Just because a link operates at a high rate does not guarantee that it shows no bufferbloat (undesired latency-increase under load), and at the same time different users/use-cases have different tolerances for bufferbloat.

2 Likes

Well, or the offload IP block runs in the same clock domain as the main CPU....

I can agree with that in most (residential) scenarios at speeds above 100mbit.

1 Like

I wrote about SQM because in many hardware selection threads many people consider SQM as a must have. That's why they overestimate hardware requirements.

1 Like

Well,IMHO that is probably just as bad advice as unconditionally declaring SQM as not necessary at speeds >= X Mbps... IMHO, all of these half-backed feature restricted off-loads are just the wrong way to deal with increasing end-user bandwidth, on the other hand I understand that simple economics will tilt the market in a way that there is no real choice for price conscious end-users...

1 Like

I wonder on which statistics do you base the "most" on?

I can explain. Here in Russia most of residential users are connected by Ethernet with symmetrical 100/100 or more links. Some are connected by optics. SQM is very unpopular here. I tried to raise this issue on some forums, but nobody is having serious bufferbloat issues. I never seen dslreports bufferbloat rate lower than B. So I never met an Openwrt user who is using SQM here.

I don't see that on 500+ megabit links bufferbloat is more of an issue.

And if you don't use SQM and don't need a very fast VPN, hardware requirements are quite different.
So we need to ask if a user needs SQM before telling that only x86 can be used for a 1000 link.

1 Like

First rule of zealotry: you start reading what you presume instead of what was written...

Again my argument is, this is not for you and me to decide whether traffic-shaping/buffer-debloating is required for any individual network short of the networks we operate ourselves. To repeat myself this is a policy decision that needs to be taken by a networks operator, and in that regard it seems important to know whether a given router will be capable of actually perform these duties at the intended bandwidths or not. Depending on a user's policy decision this can mean a (hopefully cheaper) router (with offloads) can be a cost-effective solution, or that a beefier device seems a better choice. See what I did there? I argued for supplying the information and leaving the policy decision to the local network's operator instead of declaring a priori that traffic shaping capability close to line rate either is or is not a "mandatory" requirement.

2 Likes

I believe @dlakelan operates a >= 500 Mbps internet access link that showed bufferbloat, so again it seems dangerous to extraploate from one's own limited knowledge whether traffic-shaping is required at any bandwidth X.

I wholeheartedly agree, but that decision should be left to each network's operator/admin, in the forum we should supply the necessary information to make this judgement call (and that includes assessments whether one considers this usefull), all we should refrain from is giving bad advice based on limited information.

+1

(I believer we actually agree on the core of the matter, so I will calm down again ;))

Please read again what's actually written....
"I can agree with that in most (residential) scenarios at speeds above 100mbit."
Unless you have some self-prolaimed dictatorship which the rest of us is unaware of we're all free to have our own opinion. Again, your behaviour is really tiresome regarding this matter.

You probably don't but you're still wrong :wink:

As they say, everybody is entitled to their own opinion, but not to their own facts.

Look, I agree that we disagree, but most implies >= 50% and that requires statistics or gut-feeling. I see you based that on gut-feeling, which is your prerogative. You just need to accept that a) even if only 49% of users will profit from SQM declaring it generally useless might be overly strict, and b) other's can make up gut-feely numbers using the same process that you used.
Anyway, we are not going to reach a conclusion here, so I will try to let this go.

That's exactly the thing, you think you're right even if you can't prove it. Do you have any idea how many connections any of the users in here have? I'm certain the answer is no, please just stop as it's getting really embarrasing. :wink:

Just agree that others disagree and leave it at that...

you must be tying to sell more arms and legs..

QCA9531 clocked from 550 to 1000 runs for more than 100days, placed outside in currently hot conditions, wihout any heatsink attached to cpu, no issues.
QCA9558 gets some wifi issues on very high clocks, above 1100Mhz, otherwise it is equally stable at 1000 just as QCA9531. heatsink attached to this one just becasue i had it around and it can fit in the case

Overclocking and attaching heat sinks is definitely not for a common user. :face_with_monocle:

Consumer links, even in the 300-1000 Mbps range, exhibit significant buffer bloat.

See, for example http://www.dslreports.com/speedtest/results/bufferbloat

I suspect that it will become even worse as the head-end and backhaul capacity becomes overstressed.

But then again, if you look at Com Hem Bredband, A+ A+ A+A+ for instance (Cable/Fiber provider in Sweden) so it depends...

Well, I did not start to introduce made up statistics into this discussion, so I believe I had nothing to prove. Please, put the shovel away and start to follow the first law of holes, if inside one, stop digging.

No, but I believe I made no claim that would require to estimate that number; bufferbloat is not a function of concurrent connections. Or do you claim that users operating more links (which could be construed as connections) have more say in discussions like this? Honestly, I do not know what you are arguing here.

That I will agree with and promise to let this subthread die...