Home › Topics › Firewall Monitoring Beyond Up/Down › Firewall Health Metrics

Firewall Health: The Metrics Behind the Little Green Light


There is a special kind of false confidence that comes from a monitoring dashboard where the firewall shows a friendly green light. That light usually means exactly one thing: the firewall answered a ping. It does not mean traffic is flowing. It does not mean new connections are being accepted. It does not mean both power supplies are alive. It also does not mean the box is not thirty seconds away from refusing every new session on the network. A firewall can answer pings cheerfully, right up to the moment everything stops — and sometimes even after.


Firewalls rarely fail the way up/down monitoring expects. They almost never drop dead. Instead they fail sideways. The CPU pins under inspection load and everything gets slow. The session table fills and new connections start bouncing. Memory runs out and the box quietly stops inspecting. A fan dies and the hardware starts cooking itself. Every one of those failures happens behind a green light.


This guide gives you the health metric set that actually predicts firewall trouble. It covers CPU (and why turning on inspection features multiplies it), memory, session count, interface errors, and the hardware sensors on physical boxes. You will also learn the single most useful mental model in firewall monitoring: the difference between the management plane and the data plane. And you will get a starter threshold table you can put to work this afternoon.


In this guide

  • Why the green light lies
  • The two brains: management plane vs data plane
  • CPU, inspection load, and throughput honesty
  • Memory and self-protection modes
  • Session count: the ceiling
  • Interface errors and saturation
  • Hardware health: heat, fans, power
  • A starter threshold table
  • Blamed second, cleared fast

Why the green light lies


An up/down check — a ping, usually — asks the firewall a very easy question: "are you there?" Answering it costs the firewall almost nothing. The tiny processor slice that replies to pings is one of the last things to fail, which makes ping a terrible early-warning system. It is the smoke detector that only goes off after the house has burned down.


Think about what your firewall actually does all day. It tracks every conversation crossing it and checks each one against a rule list. It rewrites addresses for NAT (network address translation — the address-swapping that lets many inside devices share outside addresses). It encrypts VPN tunnels. And — if you have turned the features on — it opens packets up to inspect what is inside them. Every one of those jobs can degrade or fail independently while the ping keeps answering. Users experience slowness, weird half-failures, and dropped calls. Your dashboard experiences serenity.


The fix is not a better ping. It is watching a small set of internal numbers the firewall will happily report — usually over SNMP, sometimes via an API or CLI. SNMP is Simple Network Management Protocol, the standard question-and-answer language monitoring tools use to ask devices about themselves. Those numbers are the rest of this guide.


The two brains: management plane vs data plane


Nearly every firewall, from small office boxes to data-center chassis, is internally split into two parts. The management plane is the part you talk to: the web interface, the SSH login, the SNMP responder, the ping reply. The data plane is the part traffic flows through: the engine that matches rules, tracks sessions, and moves packets. On bigger boxes these are literally separate processors; on small ones they are separate software processes sharing a CPU. Either way, they can fail separately — and regularly do.


This explains the classic mystery outage: the firewall answers ping and you can log into it, but no traffic is passing. The management plane is fine; the data plane is wedged. It also explains the reverse. An overloaded management plane makes the web UI crawl and SNMP time out while traffic flows perfectly. That tricks your monitoring into declaring a device "down" that users have no complaints about.


So monitor both brains, separately:


  • Management plane: ping the firewall's management address, and poll it over SNMP. If polls start timing out or slowing down, the management plane is struggling — a real signal on its own.
  • Data plane: run a check that goes through the firewall. A ping from your monitoring server (inside) to a stable target on the far side (outside) traverses the data plane the same way user traffic does. Record its response time, not just success. Rising latency through the box while direct latency to the box stays flat is an early sign the inspection engine is drowning.

If you take one habit from this article, take that pair. "Can I reach the firewall?" and "does traffic get through the firewall?" are different questions, and most firewall outages fail exactly one of them. The two answers combine into a tiny diagnosis table worth memorizing:


Answers ping/SNMP?Traffic passes?What it usually means
YesYesHealthy — or quietly struggling; check the gauges, not the light
YesNoData plane wedged or a limit hit — the classic "green light outage"
NoYesManagement plane overloaded or mis-addressed; users fine, dashboard panicking
NoNoGenuinely down — the only row an up/down check can actually see

Notice that up/down monitoring covers exactly one of the four rows. The other three are where firewalls spend most of their bad days.


CPU: where inspection multiplies the bill


Firewall CPU is not like server CPU, where a busy hour means slow reports. When a firewall runs out of CPU, it delays and then drops packets — for everyone, on every application, at once. It is the one CPU in the building with no polite way to be busy.


What burns firewall CPU? Moving packets is the cheap part. The expensive part is everything you have asked the firewall to do to those packets. Intrusion prevention (pattern-matching every packet against attack signatures), antivirus scanning of file transfers, application identification, URL filtering, and VPN encryption each add per-packet work. The heavyweight champion is TLS inspection — decrypting encrypted web traffic, examining it, and re-encrypting it. This can consume more CPU than everything else combined. Logging is not free either; a firewall logging every session works measurably harder than one logging selectively.


This is where throughput honesty comes in. The big throughput number on a firewall datasheet is measured under ideal lab conditions. Those mean large packets, simple rules, and — crucially — most inspection features switched off. Turn on the security features you actually bought the firewall for, and real-world capacity lands at a fraction of the headline figure. The datasheet usually admits this in a smaller line labelled something like "threat prevention throughput." Real traffic is also full of small packets, and per-packet costs mean many small packets hurt far more than the same bytes in large ones. None of this is vendor dishonesty exactly — every vendor measures the same way. But it is why a firewall "rated" for far more than your internet link can still run hot at lunchtime.


Tip: when a firewall feels slow, resist the urge to fix it by disabling inspection features and leaving them off. That trades a visible performance problem for an invisible security one. Measure first. If CPU is the constraint, the honest fixes are tuning what gets inspected, or admitting the box is undersized for the job you gave it.


When you monitor CPU, watch out for the averaging trap. Many firewalls have multiple cores, with some dedicated to the data plane and some to management. A dashboard "overall CPU" of 40% can hide one data-plane core pinned at 100% — and that pinned core is dropping packets. If your firewall exposes per-core or per-plane CPU (many do, over SNMP or CLI), poll those. A generic status output looks something like this:


fw01> show system status
Management CPU:     11%
Data plane CPU:     74%   (core 1: 58%  core 2: 61%  core 3: 96%  core 4: 79%)
Memory:             68% used
Sessions:           412,806 active / 1,000,000 maximum
Chassis:            temp OK, fans OK, PSU-1 OK, PSU-2 FAILED

Notice two things a green light would never tell you: core 3 is nearly saturated, and one of the two power supplies is dead. Also note the shape of a good CPU alert: sustained load, not spikes. A firewall briefly touching 95% during a backup window is normal; sitting above 80% for ten minutes is not.


Finally, learn your firewall's daily CPU rhythm before you trust any threshold. Most show a predictable curve: quiet overnight, a climb as people arrive, a lunchtime bump of streaming and updates, a tail-off in the evening. Once you know that shape, the useful question stops being "is CPU high?" and becomes "is CPU high for a Tuesday at ten?" A firewall at 60% on a Sunday night is a louder alarm than the same firewall at 75% on Monday morning. Distance from normal beats distance from zero.


Memory: the quiet number with a cliff at the end


Firewall memory holds the session table (a note per active conversation — more on that below). It holds inspection engine working space, buffered log messages, routing tables, and the signature databases those inspection features constantly consult. Unlike CPU, memory problems build slowly — which makes them perfect for monitoring and easy to sleep through without it.


Two patterns matter. The first is the slow climb: memory use creeping upward over weeks with no matching growth in traffic. That is the classic shape of a leak in some process. Left alone it ends in a crash or an unplanned reboot — on a bad day, during business hours. The second is the step change: memory jumps after a configuration change or feature enablement and plateaus at a new, higher normal. That is not a leak, but it silently shrinks your remaining headroom.


Know what your firewall does when memory runs out, before it happens. Many products enter a self-protection fallback mode under memory pressure: some refuse new sessions, some skip certain inspection work, some restart internal processes. Each behavior looks completely different from the outside — "nobody can connect" versus "traffic flows but scanning quietly stopped." Knowing which one your product chooses turns a confusing outage into a two-minute diagnosis. Your product's documentation will describe its low-memory behavior; read that paragraph now, not during the incident.


One caution on thresholds: some firewalls deliberately run at high memory use because they cache aggressively. So a flat 85% is less alarming than a climbing 70%. Alert on sustained high use, but also on trend — a steady rise over days deserves a look regardless of the absolute number.


Session count: the ceiling you will meet eventually


Every conversation through a stateful firewall occupies one entry in its session table, and that table has a hard maximum. When it fills, existing conversations keep working while every new connection fails. That half-broken state is so strange-looking that teams routinely burn hours on it. Session count deserves a graph and a threshold on every firewall you run: warn when the table passes about 70% of capacity, alert around 85–90%.


The session table is important enough — and weird enough — to earn its own guide. For the symptoms and causes, see Session Tables and Connection Limits: The Ceiling You Forgot. The causes include chatty apps, scanning malware, and sessions that never time out. The guide also covers how to hunt down the host hogging the table.


Interface errors: the box meets the wire


A firewall's interfaces fail in the same ways any network port fails, and the counters tell you which way. Input errors (CRC errors, frame errors) point outward at the physical world: a damaged cable, a failing transceiver, a duplex mismatch. Discards and overruns point inward: packets arrived intact but the firewall could not process them in time. That is a symptom of the CPU and inspection-load story above, showing up at the interface. The distinction matters because one sends you to the patch panel and the other to the performance graphs.


Watch error and discard counters on every interface that carries real traffic — inside, outside, and DMZ (the "demilitarized zone" segment where internet-facing servers live). Any steadily growing count is worth a look; the healthy value is boring, near zero, forever. And graph plain utilization on the outside interface. The firewall usually sits exactly where your fast internal network narrows to a slower internet link. So it is the natural place to notice the link itself filling up.


Hardware health: heat, fans, and the second power supply


If your firewall is a physical appliance — and at the network edge it usually is — it has a body, and bodies fail. The good news: nearly every business-grade appliance reports its vital signs over SNMP, including temperature sensors, fan speeds, and power supply status. The monitoring is easy; the discipline is remembering to set it up, because none of these produce user-visible symptoms until the very end.


The classic quiet failure is the redundant power supply. Boxes with two PSUs are bought precisely so one can die harmlessly — and it does, silently, sometimes months before anyone notices. From that moment you are running with zero redundancy while believing you have one spare. The same logic applies to fans. A dead fan changes nothing today, but the next warm afternoon takes the temperature somewhere the hardware cannot survive. Alert on any PSU or fan failure immediately; there is no such thing as a minor loss of redundancy. For temperature, baseline what "normal" looks like in your rack and alert meaningfully above it. Rising intake temperature is often really an air-conditioning alarm arriving early.


One more mundane sensor: if the firewall logs to its own disk or storage, watch free space. A full log disk ends in lost logs on every product, and on some products in genuinely strange behavior.


A starter threshold table


Every network is different, and you will tune these — but tuned-later beats absent-now. These starting points are deliberately a little conservative, because a firewall running out of anything hurts everyone at once:


MetricWarnAlertNotes
Data plane CPU70% for 10 min85% for 10 minSustained load, not spikes; watch per-core if exposed
Management CPU60% sustained80% sustainedExplains slow UI and SNMP timeouts
Memory75% used85–90% usedAlso alert on multi-day upward trend
Session table70% of capacity85–90%See the session table guide
Interface errors/discardsAny steady growthRapid growthErrors point at cables; discards at overload
TemperatureAbove your baselineNear vendor limitTrend matters; often an HVAC early warning
PSU / fan—Any failureRedundancy lost is an incident, not a note
Through-path checkLatency above baselineCheck failsTests the data plane, not the management plane

On polling frequency: every one to five minutes is plenty for all of these. Firewall health numbers move slowly enough that faster polling mostly buys you noise. The polls themselves land on the management plane. The busier you make it answering questions, the less honest its answers get. Steady polling has one useful side effect. Suppose SNMP responses themselves start arriving slowly or timing out while the box stays up. You have just measured management-plane stress without asking a single extra question.


Blamed second, cleared fast


Here is the political reality of running a firewall: it is the chokepoint everything blames second. When an application misbehaves, the first accusation is "the network is slow." When the network team proves the switches innocent, every eye swings to the firewall. That is because it touches everything and because it is the box that blocks things on purpose. It is also because nobody outside your team understands it. "The firewall must be eating it" is the second-most-popular theory in the history of troubleshooting calls.


Sometimes the theory is even right. The point of the metrics in this guide is that you can settle it in minutes either way. CPU flat, memory steady, session table at 40%, zero discards, through-path latency on baseline: "not the firewall — keep looking," said with graphs instead of vibes. Or you see the opposite: discards climbing and the data plane pinned since the moment the new inspection profile went live. In that case, you just found the problem before the meeting found you. Firewalls also get blamed for blocking specific things after rule changes. That argument is settled by logs rather than gauges, and it is covered in Firewall Logs: Alerting on What Matters.


Either way, the metrics change your role in the incident. Without them you are a suspect. With them you are a witness with evidence.


Where to go next


Start today with five checks: management-plane reachability, a through-path check, data plane CPU, memory, and session count against capacity. Add hardware sensors the same week. That is a firewall dashboard that predicts trouble instead of confirming it afterwards.


From here, go deeper on the two failure modes this guide only waved at. Explore the session table's hard ceiling in Session Tables and Connection Limits. Explore the standby unit that may be quietly dead in Monitoring Firewall HA Pairs. The rest of the pillar — rule-hit analysis and log alerting included — lives at Firewall Monitoring Beyond Up/Down.


Put this into practice with SysPeep Monitor

A firewall that answers ping can still be minutes from a bad afternoon. SysPeep Monitor reads the numbers that tell the real story — CPU, memory, session counts, interface errors, and hardware sensors, collected over SNMP. It graphs them and alerts you when they drift past thresholds you control. The free trial installs on any Windows machine in minutes.

Download Free TrialLearn About SysPeep