Capacity Planning with Cacti: Predicting the Bandwidth Cliff

In the modern era of public cloud, “capacity planning” often just means setting an auto-scaling policy and forgetting about it. If traffic spikes, the cloud provider simply spins up more instances.

But back at Net4, we lived in a different reality. We didn’t have elastic compute or burstable bandwidth on demand. We had physical circuits with hard ceilings. A 1 Gbps link was exactly 1 Gbps. If traffic hit 1.1 Gbps, the network didn’t scale; it broke. Packets dropped, latency skyrocketed, and customers went offline.

In that environment, capacity planning wasn’t an administrative task; it was a survival mechanism.

The Tool: Cacti and the SNMP Heartbeat

We relied on Cacti, a frontend for RRDTool (Round Robin Database Tool). It was the industry standard for a reason.

Every five minutes, Cacti would reach out to our border routers via SNMP (Simple Network Management Protocol) and ask a simple question: “How many octets have passed through interface GigabitEthernet0/1?”

It would plot these data points on a graph. To the untrained eye, these graphs just look like jagged mountains. To a network architect, they are a narrative of human behavior. You can see when the city wakes up, when the lunch rush hits, and when the backup jobs kick off at night.

The Analysis: Beyond the Average

The mistake most junior engineers make is looking at the “Average” utilization.

Averages are comfortable liars. If a link is idle for 12 hours and 100% saturated for 12 hours, the average is 50%. The dashboard says “Green,” but your customers are screaming for half the day.

My strategy focused on the 95th Percentile.

This metric discards the top 5% of data points (bursts) to ignore anomalies, giving you the “sustained peak” utilization—the bandwidth you are actually paying for and using.

But I went deeper. I wasn’t just checking if we were red today. I was calculating the slope.

The Bandwidth Cliff

One spring, I was auditing our primary transit link usage over a 3-month period.

On a daily basis, things looked fine. We were peaking at about 70% utilization. Plenty of headroom, right? Most teams would have closed the ticket and moved on.

But when I zoomed out to the quarterly view, I saw the trend line. Our traffic wasn’t flat; it was growing at a consistent rate of 5% week-over-week.

It was a slow creep, invisible in the daily noise. But simple math revealed the danger. Compounding that 5% growth meant that our “safe” 70% utilization would hit 100% saturation in exactly six weeks.

I looked at the calendar. My projection showed we would hit the “Bandwidth Cliff"—total saturation—by July 1st.

The Silent Win

I took the data to management. I showed them the graph, the slope, and the math.

“We are not out of bandwidth today,” I explained. “But we will be out of bandwidth on July 1st. And provisioning a new carrier circuit takes 4 weeks. We need to sign the PO today.”

We upgraded the circuit two weeks before the predicted deadline.

On July 1st, traffic surged exactly as predicted. It crossed the old 1 Gbps limit without a stutter, flowing smoothly into the new capacity. No packets dropped. No alarms triggered. No customers called.

The Lesson

In Operations, silence is the only true applause.

If you are fighting fires, you are already losing. The upgrade was a non-event, and that is exactly the point.

The lesson I carried forward is this: “Real capacity planning isn’t looking at a dashboard to see where you are today. It’s looking at the dashboard to see where you will be next month.”

The difference between a stable network and a catastrophic outage is often just a spreadsheet and the foresight to trust the trend line.