Why SQM Doesn’t Scale (And Why Datacenters Don’t Use It)

SQM (Smart Queue Management) is one of the best things that happened to home networking.

Enable CAKE or fq_codel, and suddenly:

• latency drops
• bufferbloat disappears
• the network feels “fast” again

So naturally, the question becomes:

Why don’t datacenters use it?

The answer is simple — and uncomfortable:

Because it doesn’t scale.

Why SQM Works So Well at Home

SQM solves a very specific problem:

too many flows competing for a slow link.

It works by:

• classifying flows
• ensuring fairness
• actively managing queues

On a 100 Mbps or even 1 Gbps link, this is manageable.

A modest CPU can:

• inspect packets
• maintain per-flow state
• schedule traffic intelligently

And the results are excellent.

What Changes at Scale

Now move to a datacenter:

• 25 Gbps
• 100 Gbps
• 400 Gbps

At these speeds:

• packets arrive every few nanoseconds
• flow counts explode
• memory bandwidth becomes critical

A CPU cannot:

• track every flow
• apply per-packet decisions
• maintain fairness in software

The model breaks down.

Why Hardware Wins

Datacenters don’t try to replicate SQM.

Instead, they use:

• hardware queues
• traffic classes
• simple schedulers (WRR, strict priority)

This approach is:

• less precise
• but massively scalable

And crucially:

it runs at line rate.

The Hidden Cost of “Smart” Queues

SQM is powerful because it is:

• stateful
• adaptive
• per-flow aware

But each of those has a cost:

• CPU cycles
• memory access
• cache pressure

At small scale, this is fine.

At large scale, it becomes impossible.

This Isn’t Just a Datacenter Problem

The same issue appears on small routers.

Older or low-power SoCs:

• single-core or low-frequency CPUs
• limited memory bandwidth

When you enable SQM on these devices:

• CPU becomes the bottleneck
• throughput drops
• latency improves — but at a cost

This is the same trade-off, just at a smaller scale.

The Alternative: Simpler, Aligned Systems

Instead of relying on complex software scheduling, modern designs favor:

• hardware QoS (CoS queues)
• DSCP-based classification
• end-to-end alignment (NIC, switch, Wi-Fi)

This is less flexible than SQM.

But it is:

• predictable
• efficient
• scalable

Where This Fits In Modern Routers

As we’ve seen:

• flow offload bypasses SQM
• hardware queues still work
• DSCP alignment becomes critical

This leads to a different philosophy:

don’t fight the hardware — use it.

Conclusion

SQM is not wrong.

It is just designed for a different scale.

At home, it solves real problems elegantly.

At datacenter scale — or even on small CPUs — it becomes too expensive.

That’s why large networks rely on:

• simple classification
• hardware scheduling
• consistent prioritization

Not because they are less advanced.

But because they scale.