Home › Topics › Router & WAN Link Monitoring › Router Health Metrics

Router Health: CPU, Memory, and the Forwarding Path


Your router is the busiest employee you have. Every packet that leaves the building passes through one box. Every video call, every cloud sync, every late-night backup goes through it. Most networks monitor that box with a single ping. And a ping tells you almost nothing about a router, for a reason that surprises nearly every new admin. The part of the router that answers pings is not the part that moves your traffic.


That gap explains two classic mysteries. The first: a router reports alarming CPU numbers, responds to pings slowly or not at all, and times out half its SNMP polls. Yet users notice nothing, because traffic is flowing perfectly. The second, crueler one: the router looks fine on every graph while quietly dropping your packets at an interface queue nobody was watching. Monitor the wrong metrics and you get both failure modes at once: false alarms and missed problems.


This guide fixes that. You will learn the two-machines-in-one-box mental model that makes router behavior make sense. You will learn which metrics genuinely predict trouble — CPU, memory, queue drops, temperature, power. You will learn how to set thresholds that respect what routers actually do, and why the humble small-office router deserves special attention.


In this guide

  • Two machines in one box: control plane and forwarding plane
  • The classic false alarm: high CPU, happy traffic
  • Monitoring CPU without crying wolf
  • Memory: the slow leak and the full table
  • Queue drops: the interface tells the truth
  • The hardware underneath: temperature, fans, power
  • The small-office reality check
  • A starting threshold table
  • Where to go next

Two machines in one box: control plane and forwarding plane


Inside every router live two logical machines with very different jobs. Network engineers call them the control plane and the forwarding plane (also known as the data plane). The distinction is the single most useful idea in this guide.


The control plane is the brain. It runs the router's operating system and speaks routing protocols to figure out where packets should go. It handles your SSH sessions and the web admin page, answers SNMP queries, and negotiates VPN tunnels. And — crucially for monitoring — it composes the replies when someone pings the router itself. Brain work is flexible but slow, done by a general-purpose CPU that is often surprisingly modest, because it was never meant to touch much traffic.


The forwarding plane is the conveyor belt. Its one job is moving packets from an inbound interface to the right outbound interface as fast as physics allows. It follows a lookup table the brain prepared in advance. On enterprise-class hardware the conveyor belt is built from dedicated forwarding chips that shove packets along without asking the CPU for help. It is like the way a warehouse conveyor keeps moving whether or not the shift manager is at his desk. The brain decides the rules; the belt does the work; and each can have a bad day without the other noticing.


Every router health metric belongs to one of the two planes. And every confusing router symptom becomes clear the moment you ask which plane you are actually looking at. Pings to the router, SNMP responses, config sessions: brain. Traffic through the router, interface counters, queue drops: belt. Keep those sorted and you are ahead of most of the industry's ticket queues.


The classic false alarm: high CPU, happy traffic


Now the mystery from the introduction solves itself. When a router's CPU is busy, the work being starved is brain work. Replying to your ping is one of the lowest-priority chores a router has. That is deliberate, because designers would rather the brain drop a courtesy reply than delay real duties. So a router under CPU pressure answers pings late, jittery, or not at all, and lets SNMP polls time out. Meanwhile, the conveyor belt hums along at full speed, forwarding user traffic flawlessly. Your monitoring screams; your users shrug.


This produces the most common false alarm in network monitoring: "router latency" alerts that describe the router's mood, not your traffic's experience. Routers also rate-limit their courtesy replies on principle. That is the same reason a single mid-path hop can look terrible in a traceroute while everything after it looks fine. This pattern is unpacked in localizing the slow hop.


The cure is to measure the two planes separately and compare. Probe through the router to a target on the far side — that number is the belt, and it is the one users feel. Ping the router itself — that number is the brain's responsiveness, useful as a health hint but never as a traffic measurement. When "through" is fast and "to" is slow, you have a busy brain and a working network. In that case, investigate the CPU at your leisure, and do not page anyone. When "through" degrades too, the belt itself is in trouble — usually a queue problem, coming up shortly.


Tip: build this comparison into your dashboards permanently. For each WAN router, put one chart of probe latency through it next to one chart of response time from the router itself. The pair answers "is it the router or just the router's mood?" in a single glance, and it will save you from your first three false escalations.


Monitoring CPU without crying wolf


None of this means router CPU does not matter — it means CPU is a leading indicator rather than an outage detector. A brain that is busy today can be a brain that fails to update routes tomorrow or drops a VPN negotiation. It can lock you out of the console during the exact incident you need it. The goal is to watch CPU with enough context to know which spikes are life and which are warnings.


Collect CPU over SNMP at a steady interval and treat three patterns differently. Short spikes — a minute or two of high load — are usually legitimate work. Examples include a routing update after a link flap or an SNMP walk (yes, aggressive monitoring can itself spike the brain). They include a config save or a burst of VPN negotiations after an outage. Alert on none of them. Sustained load — high CPU held for ten or fifteen minutes — deserves a warning. Routers spend most of their lives nearly idle and a new plateau means something changed. That could be new traffic being processed in software, a misbehaving process, or someone hammering the management plane. A changed baseline deserves a calendar appointment even though nothing is alerting. That includes the same router that idled at ten percent for a year now idling at forty. You have quietly spent headroom you will want during the next incident.


Where the platform exposes it, per-process CPU is the difference between knowing the brain is busy and knowing why:


router# show processes cpu sorted
CPU utilization: 91% (5 sec) / 88% (1 min) / 84% (5 min)
 PID  Runtime(ms)  5Sec   1Min   5Min  Process
  87     4418232   41%    39%    37%   Crypto Engine
  12     2190401   28%    27%    26%   IP NAT
 143      801233    9%     8%     8%   SNMP ENGINE

A readout like that turns an argument into a diagnosis. This box is spending its brain on encryption and address translation — software work that grows with traffic. The fix is capacity or offload, not a reboot. When you cannot get per-process detail, correlate the CPU graph against traffic and VPN session graphs; the culprit usually rises and falls in lockstep with its victim.


It also pays to know the usual suspects behind a busy control plane, because they narrow the search before you even log in. A flapping interface makes the brain recompute routes over and over — look for a link going up and down in the logs. Overeager monitoring is a self-inflicted classic. Polling hundreds of values from a small router every few seconds can keep its SNMP engine pinned. In that case, the kindest fix is polling less, not alerting more. And on internet-facing gear, a management plane left reachable from outside collects a steady rain of login attempts and scans. Every one is brain work. Blocking that traffic upstream often returns a startling amount of CPU.


Memory: the slow leak and the full table


Router memory problems come in two shapes, and neither is subtle on a graph. The first is the slow leak: free memory declining week over week, staircase-style, without ever recovering. Some process is allocating and never releasing. And the endgame is a router that refuses new routes, drops VPN tunnels, or reboots itself at an hour chosen by fate. Memory leaks are gloriously easy to catch with monitoring — the graph is a straight line pointing at a date. They are nearly impossible to catch without it, because nothing feels wrong until the cliff edge. If the trend line says the box runs dry in six weeks, you get to schedule the fix instead of attending the outage. The fix is usually a software update or a planned reload.


The second shape is the full table. Routers keep several finite tables in memory — the routing table and the ARP table. Any box doing NAT or stateful firewalling also has a connection table with one entry per active session. Tables fill under unusual load: a malware-infected PC opening thousands of connections a minute is the classic. When the connection table fills, the router starts refusing new sessions while existing ones work fine. That is a maddening symptom pattern ("old tabs work, new tabs spin") that makes no sense until you look at the table gauge. Monitor table usage wherever the platform exposes it, and alert well before full. The alert usually points at a sick client on your LAN, not at the router at all.


Thresholds for plain memory: warn when free memory falls below roughly a quarter and hold there. Alert below ten percent. Separately, watch the trend — a healthy router's memory graph is boring and flat. Any steady slope is worth a ticket regardless of the current level.


Queue drops: the interface tells the truth


Here is the metric that catches the problems CPU and memory graphs miss — the one that speaks for the forwarding plane itself. When packets arrive for an outbound interface faster than the line can carry them, the router buffers them briefly in a queue. When the queue is full, it throws packets away and increments a counter: output drops (reported as output discards or output queue drops, depending on the platform). That counter is the forwarding plane's confession, and unlike almost everything else on the box, it never exaggerates and never speaks of moods. Every increment is a real packet of yours that died.


The reason this counter is precious is that utilization graphs lie by averaging. A five-minute average of forty percent looks innocent, but traffic is bursty on scales of milliseconds. A microburst can slam an interface to line rate for half a second, overflow the queue, and vanish, leaving the average unbothered. Users feel a stutter on every call; the utilization graph swears nothing happened; the drop counter tells the truth. If you take one habit from this guide, make it this. On every WAN-facing interface, monitor output drops and input errors alongside utilization. Treat a climbing drop counter as real congestion no matter how relaxed the utilization graph looks. (Input errors are a different confession — packets that arrived damaged, pointing at cables and physical trouble on the line. The circuit-side reading of those counters is covered in WAN circuit health.)


Interface WAN1   5-min average utilization: 41%
  Output queue: 0/40 (current/max)
  Total output drops: 18,449     <-- climbing every business day
Interface LAN1   5-min average utilization: 12%
  Total output drops: 0

Read as a pair, those two interfaces tell a complete story. The LAN side is loafing. The WAN side is microbursting into a full queue at rush hour. And any "the internet stutters" complaints are congestion at the WAN queue — invisible to averages, undeniable in drops.


The hardware underneath: temperature, fans, power


Below both planes sits the part of the router that obeys thermodynamics. Most business-class routers expose environmental sensors over SNMP, and they are cheap insurance to collect. Temperature is the big one. A router in a stuffy closet cooks slowly. Overheating hardware fails in the least diagnosable ways — random reboots, flapping interfaces, corrupted forwarding behavior that no config change explains. A temperature graph that trends upward over weeks is usually announcing a dying fan or a failing closet air conditioner. Both are far cheaper to fix before the crash. Alert when the sensor approaches the top of the operating range the vendor publishes for the platform, and warn well below it.


Fan status is binary and blunt: a failed fan on a passive shelf is a countdown timer. Power supply state matters most on hardware with two supplies. The whole point of dual PSUs is that one can die without an outage. That also means one will die without an outage. Without monitoring, nobody finds out until its partner follows. A dead redundant supply is the quietest alert class in networking and one of the most valuable. The fix is a part swap on a calm Tuesday instead of a dead router on a frantic Monday. (You will meet this exact pattern again at circuit scale — redundancy silently lost while service hums along — in multi-WAN visibility.)


One more hardware-adjacent check earns its keep: uptime. Routers reboot themselves sometimes — a power blip, a crash, a watchdog reset. A box that goes down at 3 AM and returns ninety seconds later leaves no trace a morning ping will find. The router's uptime counter (exposed over SNMP as sysUptime) makes the invisible visible: if the value ever goes down between polls, the box restarted, full stop. Alert on that reset. A router that silently reboots once is a mystery; one that does it weekly is a failing power feed, a heat problem, or a crash pattern. The uptime graph is the only witness.


The small-office reality check


Everything so far applies doubly to the box guarding most small networks. It is the all-in-one edge device that is simultaneously router, NAT gateway, firewall, VPN concentrator, DHCP server, and sometimes the Wi-Fi controller. On these boxes the tidy two-plane story bends, because there is no hardware conveyor belt — a general-purpose CPU plays both brain and belt. Much of the traffic path runs in software: every NAT translation, every firewall rule evaluation, every byte of VPN encryption is CPU work.


Three practical consequences. First, on an all-in-one box, high CPU is not automatically a false alarm. The belt really can starve, so pair every CPU alert with the through-probe before deciding how loudly to react. Second, throughput claims on the spec sheet rarely survive contact with features. A box that forwards at an impressive rate with everything disabled may manage a fraction of it with firewalling and VPN encryption running. The gap is pure CPU. If users complain the moment anyone starts a big transfer over the VPN, graph CPU against VPN traffic and watch them move together. Third, the connection table from the memory section is front and center here. NAT means every session occupies an entry — monitor it on any all-in-one box that exposes it.


Baselines matter more than judgments borrowed from bigger hardware. An enterprise router idling at sixty percent CPU is alarming. A small-office box that has cheerfully run at sixty percent every business day for a year is just being itself. And the useful alert is "unlike itself," not "above a number someone posted on a forum."


A starting threshold table


Numbers you can start with today, tuned to your own baselines as they accumulate. All of these come from the router over SNMP (or SSH where SNMP does not expose a value).


MetricPlaneWarnAlertNotes
CPU loadControl>75% for 10 min>90% for 15 minIgnore short spikes; watch baseline shifts
Free memoryControl<25%<10%Any steady downward trend = ticket
Connection/NAT tableBoth>70% full>90% fullSpike usually means a sick LAN client
Output drops (WAN)ForwardingAny sustained climbRapid climb in business hoursCongestion truth; trust it over averages
Input errors (WAN)ForwardingAny sustained climb—Physical-layer symptom; check cabling first
TemperatureHardwareNearing vendor limitAt vendor limitTrend up over weeks = cooling problem
Fan / PSU statusHardware—Any failureDead redundant PSU is still an alert
Probe through routerForwardingLatency above baselineLoss above baselineThe user-experience tiebreaker

Treat the table as scaffolding. Run it for a few weeks, note which alerts fired and whether you cared, and adjust until every alert is one you would act on. A threshold you routinely ignore is not protection — it is training in ignoring alerts.


Where to go next


You now have the router's side of the story: brain, belt, and the hardware they stand on. The other half of WAN health is the line itself — WAN circuit health covers the five metrics that describe a circuit and the baseline register every line deserves. From there, the rest of the Router & WAN Link Monitoring pillar builds outward. It covers watching two circuits as one service and proving failover works before you need it. It also covers monitoring links to sites where nobody is home.


Put this into practice with SysPeep Monitor

Routers will happily report every metric in this guide — CPU, memory, drops, temperature, power — if something asks. SysPeep Monitor asks: it discovers your routers and interfaces automatically. It collects health metrics over SNMP and SSH. It lets you set warn and alert thresholds with dependencies that keep one busy box from paging you five times. The free trial runs on any Windows machine.

Download Free TrialLearn About SysPeep