Home › Topics › The Core Network Metrics › Metric Misreads

Eight Classic Misreadings of Network Metrics


A monitoring system never lies to you on purpose. But between the wire and your eyeballs sit polling intervals, averages, counter arithmetic, autoscaling axes, and protocol quirks. Every one of them can make a healthy network look sick, or a sick one look healthy. Sooner or later, every admin acts on a graph that was technically telling the truth… about something other than what they thought.


This guide collects the eight misreadings that catch almost everyone, junior and senior alike. For each one you get the misread itself, and why the illusion happens. You get a short story of how it plays out in real life, and the reading technique that gets you to the truth. You will not need any new tools by the end — only new questions. And because these mistakes are so common, knowing them has a second payoff. You will start spotting them in other people's conclusions, politely, before anyone emails the ISP in anger.


In this guide

  • 1. The calm graph that hid the storm
  • 2. The router that "drops 30% of packets"
  • 3. The 40-gigabit spike on a gigabit port
  • 4. The Alps that were a lawn
  • 5. The firewall "pegged at 100% CPU"
  • 6. Measuring through the bottleneck
  • 7. Round-trip numbers, one-way problems
  • 8. "The uplink is at 20%, so there is headroom"
  • The eight at a glance
  • The healthy-skepticism checklist

1. The calm graph that hid the storm


The misread: users report freezes, choppy calls, or momentary stalls. You check the link's utilization graph and it cruises along at a polite 35%. Verdict: "the link is fine, it must be something else."


Why it happens: most bandwidth graphs are built by reading a device's traffic counters on a schedule — commonly every five minutes. Each point on the graph is the average rate across that whole interval. Averages are ruthless flatteners. A microburst — a few seconds of total saturation — drowns in the surrounding calm:

15 seconds at 100%  +  285 seconds at 20%
= (15 x 100 + 285 x 20) / 300
= 24% average for that five-minute point

The users lived through fifteen seconds of a completely full link — queues overflowing, packets dropping — and the graph faithfully reports a serene 24%.


Seen in the wild: every evening when the backup job kicks off, phone calls go robotic for about a minute. The bandwidth graph shows a gentle hill topping out near 40%, so the network team keeps blaming the phone vendor. In reality the backup's opening burst saturates the uplink until its transfer settles down. The five-minute average never comes close to showing it.


The correct reading: first, always know the polling interval behind a graph — an averaged graph simply cannot show events shorter than its interval. Second, look at the counters that accumulate what averages hide: interface drop and discard counters climb during every microburst and never average away. Third, use peak or maximum rollups and 95th-percentile views where your tool offers them. And during an active investigation, temporarily poll the suspect link every 10–30 seconds; the storm appears immediately.


2. The router that "drops 30% of packets"


The misread: a traceroute-style tool such as mtr or pathping shows heavy packet loss at hop 2 or 3 — usually an ISP router. Verdict: "found it, the ISP is dropping a third of our traffic."


Why it happens: a router's real job — forwarding your packets — runs in dedicated, very fast hardware. Answering probes addressed to the router itself is a different activity. It is handled by the device's small management CPU, as its lowest-priority chore, and is often deliberately rate-limited. So the router forwards your traffic flawlessly while snubbing your probes. The tool honestly reports that its probes went unanswered; the misread is concluding that user traffic shares their fate.

HOST: admin-ws           Loss%   Snt   Avg
  1.|-- 10.0.0.1          0.0%   100   0.5
  2.|-- 203.0.113.1      30.0%   100   9.1   <- scary, but…
  3.|-- 198.51.100.9      0.0%   100   9.4
  4.|-- 192.0.2.50        0.0%   100   9.6   <- …the truth is here

Seen in the wild: a junior admin attaches the screenshot to a strongly worded ISP ticket. Support replies, with the weariness of long practice, that the final hop shows 0.0% loss and no fault is found. It is a rite of passage; nearly everyone files this ticket exactly once.


The correct reading: loss is real only when it starts at some hop and then persists through every hop after it, including the destination. That pattern means probes crossing that segment genuinely die. Loss that appears at one middle hop and vanishes afterward is the router deprioritizing you, not dropping customers. Read the bottom line first; it is the end-to-end truth.


3. The 40-gigabit spike on a gigabit port


The misread: the bandwidth graph shows a monstrous spike — tens of gigabits per second on a one-gigabit port — or occasionally a negative dip. Verdict: "we are being attacked" or "the graphing is broken, ignore it all."


Why it happens: traffic counters are running totals of bytes handled since boot. Older devices keep them in 32-bit counters, which can only count up to 232 bytes before rolling over to zero like a car odometer. Your monitoring computes rate as (new reading − old reading) ÷ time. When the counter wraps between polls — possibly more than once — that subtraction produces nonsense:

2^32 bytes           = 4,294,967,296 bytes (about 4.3 GB)
1 Gbit/s             = 125,000,000 bytes per second
time to wrap 32 bits = 4,294,967,296 / 125,000,000 = about 34 seconds

Thirty-four seconds — against a five-minute polling interval. A busy gigabit link can lap the odometer eight times between polls, and no software can reconstruct how many laps it missed.


Seen in the wild: the monthly traffic report features a 12 Gbit/s spike on the WAN at 2 AM. Management wants to know whether the company was hacked. The answer: the nightly backup wrapped a 32-bit counter several times, and the graphing dutifully plotted the arithmetic.


The correct reading: apply the physics ceiling — no honest reading can exceed the link's actual speed, ever. So any point above it is an artifact by definition. Then fix the source: poll the 64-bit high-capacity counters (ifHCInOctets and friends), which effectively never wrap. If a device only offers 32-bit counters, poll it faster than its fastest possible wrap time. The SNMP Monitoring guides cover these counters from the ground up.


4. The Alps that were a lawn


The misread: someone posts a latency graph that looks like a mountain range — jagged peaks, plunging valleys — and the room concludes the network is unstable.


Why it happens: dashboards autoscale the vertical axis to fit the data, stretching whatever range exists until it fills the frame. If latency wandered between 9.0 and 9.6 milliseconds all day, the graph draws that 0.6 ms of wander as cliffs. Autoscaling is a reasonable default for spotting change. It is a terrible default for judging magnitude, because it gives a flat, healthy metric the same dramatic silhouette as a genuine incident.


Seen in the wild: a screenshot lands in the team chat: "look at these latency spikes!!" Ten minutes of concern later, someone finally reads the y-axis: it spans 9.0 to 9.6 ms. The terrifying summit is six-tenths of a millisecond tall — smaller than the measurement noise. Meeting adjourned, slightly sheepishly.


The correct reading: numbers before shapes. The first glance at any graph goes to the axis range and units; the silhouette gets read second, and only in that context. For dashboards other people see, pin the scales — zero-based or a fixed sensible range — so today's graph is visually comparable to yesterday's. And before reacting to any dramatic shape, ask the disarming question: would this survive on a zero-based axis? Most Alps do not.


5. The firewall "pegged at 100% CPU"


The misread: the firewall's CPU graph spikes to 95–100% at regular intervals. Verdict: "the box is overloaded — we need a bigger one," followed by a budget request.


Why it happens: a CPU value fetched over SNMP is a gauge — a right-now number. On many devices it reflects a single instant or a very short window. Network devices also run periodic housekeeping: statistics generation, configuration backups, route refreshes, and, pleasingly, the work of answering your monitoring poll. If the sampled instant lands on a housekeeping beat, you record 100%, even though the other 59 seconds of that minute were idle. Many devices expose several flavors of the metric — instantaneous, one-minute average, five-minute average — and polling the instantaneous one manufactures drama.


Seen in the wild: a firewall spikes to 100% exactly on the hour, every hour, and a replacement is proposed. Investigation shows the spike coincides with the device's own scheduled report job; user traffic sails through untouched. Meanwhile the genuinely worrying number was a sustained 85% through the lunch peak, climbing month over month. That had been ignored, because the spikes hogged all the attention.


The correct reading: separate instantaneous from sustained. Act on values that stay high across several consecutive polls — minutes, not moments. Prefer the device's own averaged CPU values where offered. Correlate with symptoms: did throughput dip, did sessions drop, did anyone actually feel anything at spike time? A spike with no symptom is a note in your log. A sustained plateau with symptoms is a capacity conversation. Only one of them belongs in a budget request.


6. Measuring through the bottleneck


The misread: "the office internet is slow — I tested it myself." The test ran from a laptop on hotel Wi-Fi, across a VPN, to a speed-test site. Or the office's shiny new gigabit uplink gets condemned as "40 megabits at best" — measured from a laptop two walls away from the access point.


Why it happens: every measurement includes everything between its two endpoints, and reports roughly the capacity of the narrowest link in that chain. The hotel tester's path crossed laptop Wi-Fi, hotel broadband, VPN encryption overhead, the VPN concentrator, and only then the office link. That is five suspects, one number, and the number belongs to the weakest of them. It is weighing a whole chain and announcing the weight of one link.


Seen in the wild: an executive on conference-season hotel Wi-Fi reports that headquarters' internet "is down to 8 megabits." On-site wired tests show the full contracted rate within the hour — but the afternoon still goes to writing a diplomatic explanation of hotel Wi-Fi.


The correct reading: before trusting any measurement, list every link between its two endpoints — actually write the path down. Then move the instrument next to the thing being measured. Judge the office uplink from a wired office machine. Judge the server path from the server's own network segment. A test that crosses Wi-Fi or a VPN is a test of that Wi-Fi or VPN, whatever it says on the label.


7. Round-trip numbers, one-way problems


The misread: users say uploads crawl or their calls break up, but ping times look normal-ish and download tests are great. Verdict: "the network is fine; must be the app."


Why it happens: a ping time is a round trip — the sum of the outbound trip and the return trip, wearing a single number. A 100 ms RTT could be 50 out and 50 back, or 95 out and 5 back; ping cannot tell you. Yet many real problems live in one direction only. The classic is a saturated upstream. Some heavy upload fills the outbound queue. Everything leaving the building waits in that queue, and outbound delay balloons. Meanwhile the inbound direction stays pristine and every download-flavored test comes back glowing.


Seen in the wild: video calls disintegrate every afternoon. Download speed tests are spotless, so the ticket gets closed twice. The eventual culprit was a workstation backup client saturating the modest upstream every afternoon. Everyone's outgoing audio was queueing behind one machine's backup.


The correct reading: test each direction on purpose. With iperf3, run the test normally and again with -R (reverse) and compare. Run a continuous ping while a large upload runs. If latency climbs from 12 ms to 300 ms, you have just watched the outbound queue fill in real time. And read interface graphs per direction. Inbound and outbound utilization are two separate numbers, and averaging them in your head hides exactly this failure.


8. "The uplink is at 20%, so there is headroom"


The misread: the monitored WAN or core uplink shows comfortable utilization. So the complaint about slow transfers gets dismissed: "can't be the network, the link is barely used."


Why it happens: utilization is a fact about one interface. A user's traffic crosses many: their access port, an aging closet switch, the firewall, the ISP's shaping, the far server's connection. End-to-end capacity is the minimum across every link on the path — the convoy moves at the speed of its slowest truck. Monitoring tends to watch the impressive core links, while the narrow spots hide at unmonitored edges. So the one graph you have is healthy, and the bottleneck is somewhere you have no graph at all.


Seen in the wild: a branch office's file transfers crawl for a week while the head-office WAN graph shows 25% and the ISP absorbs the blame. Then someone finally checks the branch switch itself. After a cable swap, the uplink port auto-negotiated down to 100 Mbit/s half duplex. Its error counters were screaming into the void. Nothing on the monitored WAN graph ever looked wrong, because nothing on that interface ever was.


The correct reading: enumerate the actual path, hop by hop, with each link's speed — a two-minute table that instantly reveals the narrowest point. Verify end-to-end with a real transfer test between the endpoints that matter instead of inferring from one interface's graph. And when complaints cluster somewhere, extend monitoring to the edge ports there, or at least spot-check their speed, duplex, and error counters. The periphery is where this misread breeds.


The eight at a glance


#The misreadTell-tale signFirst question to ask
1Averages hiding spikesReal complaints, calm graphsCould the event be shorter than the polling interval?
2"Loss at the router"Mid-path loss, clean final hopDoes the loss persist to the destination?
3Counter-wrap spikesReadings above the link's physical speedAre these 32-bit or 64-bit counters?
4Autoscaled dramaWild shape, tiny axis rangeWhat does the y-axis actually span?
5CPU spike panicPerfectly regular spikes, no symptomsInstantaneous sample, or sustained average?
6Testing through the bottleneckTest crossed Wi-Fi or a VPNWhat sits between the two endpoints?
7Round-trip blindnessDownloads fine, users still sufferingWas each direction tested separately?
8One interface = the pathHealthy core graph, edge complaintsWhat is the slowest link on the actual path?

The healthy-skepticism checklist


Before acting on any surprising graph or alarming number — and especially before forwarding it to anyone — walk this list:


  • Read the y-axis first. Range and units before silhouette.
  • Ask the polling interval. Could the event be shorter than one graph point?
  • Identify the flavor. Average, peak, or instantaneous sample — which one is this?
  • Apply the physics ceiling. Any value above the link's actual speed is an artifact, full stop.
  • For loss: check the final hop. Only loss that persists to the destination is end-to-end loss.
  • Name both endpoints. What exactly was measured, from where to where, and what sits in between?
  • Check both directions. A round-trip number is two one-way stories.
  • Correlate with humans. Did anyone actually feel anything when the graph looks worst?
  • Get a second instrument. Does an independent tool, measuring differently, agree?
  • Ask what changed. A config, a cable, a polling setting — artifacts have birthdays too.

Tip: a surprising graph is a hypothesis, not a verdict. The most useful sentence in a network investigation is "before we act on this — let's confirm it with a second measurement." It costs five minutes and it has saved countless ISP tickets, hardware purchases, and afternoons.


Where to go next


Every misread on this list is really the same lesson wearing eight costumes. A metric is an instrument reading, and instruments have quirks you must know before trusting them. These include sampling windows, counter sizes, direction blindness, and the position of the probe. Learn the quirks and the same graphs that fooled you become genuinely trustworthy tools. The other guides in The Core Network Metrics build that fluency from the other side. They cover what each metric means, what healthy values look like, and how to measure them yourself with free tools.


Put this into practice with SysPeep Monitor

Good tooling will not read graphs for you, but it can remove the classic traps. SysPeep Monitor collects interface and health metrics over SNMP and WMI. Its threshold and dependency rules are designed to cut alert noise. So the surprises that do reach you are likelier to be real. You still bring the healthy skepticism — the software just gives you cleaner numbers to be skeptical about. The free trial installs on any Windows machine in minutes.

Download Free TrialLearn About SysPeep