Title: Congestion Control breaks under real load – CPU overload on R3 (hardware limitation?)

[Sergejs Kotovs]

I’m creating a new thread to clearly summarize the real issue, based on extensive real-world testing.

My internet connection itself is stable. This is not an ISP or line problem.
Bandwidth is set manually and correctly.

The core problem is how Congestion Control and Bufferbloat behave under real load.

When I reduce bandwidth slightly (for example a small reduction of ~5%), the router behaves as if the bandwidth was reduced massively, sometimes closer to cutting it in half.
The reduction is not proportional to the value set.

At the same time, bufferbloat testing becomes unreliable:

• throughput shown in bufferbloat tests is far lower than expected

• latency spikes heavily under load

• results fluctuate and look unstable / glitchy

• real traffic causes spikes, stutter, and lag

In other words:
small bandwidth changes cause disproportionately large throughput drops, while queue control still fails.

Important observations:

• The built-in speedtest inside the R3 interface is always stable and consistent

• When QoS / speedtest bypass is enabled, external speedtests immediately become stable and show correct speeds

• With other routers on the same connection, external speedtests and bufferbloat behave normally and consistently

This strongly suggests that under real traffic:

• the R3 CPU becomes overloaded

• once the CPU is saturated, Congestion Control no longer works correctly

• shaping stops scaling properly

• bufferbloat testing no longer reflects reality

This does not look like a tuning issue.
All common adjustments have already been tested extensively.

The behavior points to a hardware / CPU performance limitation, where the router cannot reliably handle real-world traffic shaping and bufferbloat control at this connection speed.

My questions are:

1. Is this non-linear bandwidth behavior and bufferbloat instability under load expected when the R3 CPU is saturated?

2. Is there any real fix or workaround, or is disabling shaping the only realistic option?

3. If this is a hardware limitation, can this be confirmed clearly?

4. The screenshots I attached were taken during active bufferbloat testing.

[Krush]

Pas besoin de multiplier les sujets STP ! 
C’est un comportement normal lorsque tu effectues ce type de test avec la page home ouverte de la GUI. Elle consomme plus de charge CPU.

C’est comme effectuer un benchmark de GPU avec un jeu en cours sur un PC …

Tu peux effectuer le même test sur un autre onglet de la GUI et nous dire comment se comporte la charge CPU stp 😉

[Netduma Fraser]

No need to create a new topic, I'll close this one and respond to your first topic.