Endgame Bufferbloat Results: RT-AX86U Pro + CAKE SQM (+0ms Active Latency)

[Viktor Jaep]

Has anyone else replicated +0ms latency under load with these settings and commands?

I never knew there was such a thing as 0ms latency, unless you were plugged in directly to a switch where the server would be.

 

[dave14305]

I never knew there was such a thing as 0ms latency, unless you were plugged in directly to a switch where the server would be.

Well, "incremental" latency while under load. Are these tweaks repeatable by anyone else?

But it also raises another question: were these benchmarks run over ethernet or wireless?

 

[Fonpap]

Well, "incremental" latency while under load. Are these tweaks repeatable by anyone else?

But it also raises another question: were these benchmarks run over ethernet or wireless?

Hi Dave,

To answer your questions:

Well, "incremental" latency while under load. Are these tweaks repeatable by anyone else?

But it also raises another question: were these benchmarks run over ethernet or wireless?

Hi Dave,

To answer your questions:

1. Ethernet or Wireless? These benchmarks were strictly run over Ethernet. Achieving "Lab Perfect" results with +0ms incremental latency under full load is practically impossible over Wi-Fi due to inherent jitter and airtime contention.
2. Repeatability & CPU Core Distribution: Yes, these tweaks are repeatable, but there is a crucial detail many overlook. If you want to see these results, you must monitor your router's System Status (Network Map) during a Bufferbloat test.
   
   • If you see Core 1 hitting 100% while the others are idling, you will never get these results. The "bottleneck" isn't your line; it's the CPU's inability to handle the interrupts efficiently.
   • My script uses IRQ Steering / smp_affinity and RPS (Receive Packet Steering) to force the network load to be distributed across all CPU cores. When the load is balanced and the CPU has "headroom" (90%+ idle), the CAKE algorithm can manage the queues with microsecond precision, resulting in that +0ms spike.

The "secret sauce" is ensuring the router's CPU isn't choking on a single core while processing the SQM.

 

[anotherengineer]

Has anyone else replicated +0ms latency under load with these settings and commands?

If I knew how to script and ssh and stuff I would give it a go for something to do. Alas I don’t know how. Also I’m not pppoe, I think it’s RFC 1483 protocol iirc?

 

[aru]

Could you share your results from https://bufferbloat.libreqos.com/vh/ ?

I'm on a 500M/500M symmetrical connection. Here are my baseline results with CAKE QoS disabled:

Curious whether applying your config would make any meaningful difference?

 

[Fonpap]

By the way, how did you get the 'Household Load Timeline' graph to display? I'm curious because on my end, since my latency stays perfectly flat at +0ms / +0ms, Waveform just shows the Box Plots (cluster of dots) at the baseline instead of a timeline. Did you use a specific browser or a different test mode to see the full graph even with spikes?

 

[Fonpap]

Wait, I just saw the LibreQoS link at the top of your post! I was looking for that timeline on the Waveform site, that's why I was confused. Checking LibreQoS now!

 

[Tech9]

my latency stays perfectly flat at +0ms / +0ms

Like this?

This is Bandwidth Limiter only to 250/50 on my "jittery" DOCSIS line.
On FTTH line it is going to be the same with <5ms Unloaded latency.

 

[Tech9]

Want it faster?

Bandwidth Limiter 400/50, with network traffic and via AnyDesk from 7000km away:

Hardware: $130 UCG-Ultra, CPU: Qualcomm IPQ5322 quad-core ARM Cortex-A53 1.5GHz.

 

[Tech9]

Wait, I just saw the LibreQoS link at the top of your post!

Want this one and even faster?

Reminder - no QoS, just Bandwidth Limiter, this one at 500/50. Where is the Bufferbloat?

 

[Fonpap]

Want it faster?

Bandwidth Limiter 400/50, with network traffic and via AnyDesk from 7000km away:

View attachment 71047

Hardware: $130 UCG-Ultra, CPU: Qualcomm IPQ5322 quad-core ARM Cortex-A53 1.5GHz.

Nice hardware with the UCG-Ultra, but 21ms base latency is a bit high for a 'fast' setup.


Here are my results using an RT-AX86U Pro with custom core balancing (RPS pinning) and CAKE. As you can see, I’m pushing 537/264 Mbps—which is significantly higher bandwidth than 400/50—while maintaining a rock-solid +0ms / +0ms bufferbloat grade.


The real proof is in the consistency: My average ping is 6.0ms with a standard deviation (stddev) of only 0.290ms over a 500-packet sample while the line is active.


To see how that Qualcomm CPU actually handles the stress, can you show us your ping statistics (min/avg/max/stddev) for 500 packets (ping 1.1.1.1) specifically while the bufferbloat test is running? That’s where the real stability of the hardware is proven. If it’s flat and clean — don’t touch it.

 

[Tech9]

but 21ms base latency is a bit high

Nothing I can do about it, DOCSIS line. On FTTH lines I have <5ms and +0/+0 on Upload and Download. Asking one more time - Where is the Bufferbloat below line saturation?

can you show us your ping statistics

It's constant ~18ms on DOCSIS and ~2ms on FTTH. I have local Cloudflare server in the city. I think what I showed already covers excellent quality gameplay as per Warzone bandwidth and latency requirements (50-100Mbps, 0-30ms) without applying any form of QoS. Exactly what I told you in the beginning of this thread - if you stay below line saturation and your ISP line is good quality there is nothing much to optimize. You have limited QoS control on your end and will rather hurt yourself then fix something.

 

[Fonpap]

Nothing I can do about it, DOCSIS line. On FTTH lines I have <5ms and +0/+0 on Upload and Download. Asking one more time - Where is the Bufferbloat below line saturation?



It's constant ~18ms on DOCSIS and ~2ms on FTTH. I have local Cloudflare server in the city. I think what I showed already covers excellent quality gameplay as per Warzone bandwidth and latency requirements (50-100Mbps, 0-30ms) without applying any form of QoS. Exactly what I told you in the beginning of this thread - if you stay below line saturation and your ISP line is good quality there is nothing much to optimize. You have limited QoS control on your end and will rather hurt yourself then fix something.

I agree that staying below line saturation helps, but that's exactly why SQM exists. A simple Bandwidth Limiter is a static 'dumb' tool, whereas CAKE is dynamic and flow-aware. It doesn't just prevent bloat; it ensures that even if a background process (like a system update or a phone backup) starts during a match, your gaming packets won't see a single millisecond of jitter.

 

[dave14305]

@Fonpap Will you fix the code snippet in post #56? It's got html tags in it from your copy/paste.

 

[Tech9]

it ensures that even if a background process (like a system update or a phone backup) starts

Sorry, I did the tests with some unknown to me traffic in background. You can see the traffic spike where the LibreQoS test was done. You don't see the previous Waveform tests on this graph because they were done on different ISP line. I have 4x gateways* on different ISPs and +0ms/+0ms on Upload and Download is easily achievable if the line bandwidth is not stressed. I understand your theory, but it doesn't automatically translate to reality for everyone, especially on FTTH connections.

* - one is in your country on Cosmote