Home › Topics › Finding Physical-Layer Problems › Error Counter Field Guide

Interface Error Counters: A Field Guide


Every switch port keeps a diary. It counts every frame that arrived damaged, every fragment of a collision, every time the link blinked out and came back. It is a complete confession of everything the physical path has done wrong, sitting there free of charge. And most of it goes unread. The typical admin glances at "errors: 0," nods, and moves on. The typical unlucky admin glances at "errors: 2,847,119," and feels a chill. That admin still isn't sure what it means or whether it happened today or last spring.


This guide makes you fluent in the diary. You'll learn what each of the important counters physically means — in plain language, not standards-committee prose. You'll learn how to read them over time so the numbers tell you when and how fast. You'll learn how to tell a slowly failing cable from a bursty source of interference. And you'll learn how to walk a real diagnosis from counters alone. By the end, a screen of interface statistics will read less like hieroglyphics and more like a witness statement.


In this guide

  • Where the counters live, and their one big quirk
  • The counters, one by one
  • The quick-reference table
  • Totals lie: read deltas instead
  • Steady trickle or bursts: two different diseases
  • A worked diagnosis
  • Polling and alerting in practice
  • Where to go next

Where the counters live, and their one big quirk


Every managed switch, router, and server network card maintains these numbers per interface. You can read them two ways. Interactively, from the device's command line — some flavor of a "show interfaces" command prints them all in a block. Or read them programmatically, over SNMP (Simple Network Management Protocol), the standard question-and-answer protocol monitoring tools use. The general error counts live in the standard interface table under names like ifInErrors and ifOutErrors. The detailed Ethernet-specific ones (CRC, collisions, symbol errors) live in an Ethernet statistics group most vendors expose. If SNMP is new to you, the SNMP monitoring guides cover it from zero.


The one big quirk you must internalize: these counters are cumulative. They start at zero when the device boots (or when someone clears them) and only ever climb. The number on your screen is a lifetime total, not a current rate. A port showing 94,000 errors might be in agony right now — or might have survived one terrible afternoon long ago and been an angel ever since. The total alone cannot tell you which. Everything in the second half of this guide flows from that single fact.


A smaller quirk worth knowing: counters have a maximum. Older 32-bit counters wrap around to zero after about 4.3 billion, which busy traffic counters can hit surprisingly fast on gigabit links. Error counters wrap far more rarely, but a monitoring system polling frequently handles wraps gracefully either way. Prefer the 64-bit versions where your gear offers them. Don't let a counter that seems to have gone backward spook you — it wrapped, or someone cleared it.


The counters, one by one


Input errors and output errors: the umbrella counters

Most platforms roll their detailed counters up into two totals. Input errors is the umbrella over everything that arrived broken — CRC failures, runts, and friends. It's the number to glance at first, because any nonzero growth there sends you into the detailed counters below to learn which kind of broken. Output errors count frames the port failed to transmit properly. They carry a different accusation. Damage on the way out usually means the device's own transmit hardware or its immediate connector is struggling, since nothing else has touched the frame yet. Input errors blame the path and the far end; output errors blame the port you're standing on. That asymmetry is a free localization clue on every screenful.


CRC / FCS errors: the frame arrived, but damaged

Every Ethernet frame ends with a frame check sequence (FCS) — a checksum computed by the sender over the whole frame. The receiver redoes the math (a cyclic redundancy check, CRC). If the answers differ, at least one bit changed in transit, and the frame is discarded and counted here. Physically, a CRC error means the signal got hurt between the two ports. Causes include electrical noise leaking into a copper pair, a weak or dirty optical signal, a corroded contact, or a failing transceiver. Or the frame was chopped mid-flight by a collision, which is why duplex mismatches also feed this counter. This is the headline counter for physical diagnosis. A modern full-duplex link in good health should show essentially none.


Alignment errors: the frame didn't end where bytes end

Data travels in 8-bit bytes, so a valid frame's length is always a whole number of bytes. An alignment error is a frame that arrived with a ragged edge — some leftover bits — and a bad checksum. Something interrupted or mangled the transmission mid-byte: classically a collision on an old half-duplex link, otherwise serious noise. On modern full-duplex networks this counter mostly appears alongside CRC errors and points at the same suspects. Climbing alignment errors with collisions present is a strong hint of duplex trouble.


Runts: frames too small to be legal

Ethernet's rules say a frame must be at least 64 bytes; anything shorter is a runt. Runts are usually not tiny messages — they're stumps: the front end of a frame whose transmission was aborted partway, most often by a collision. A trickle of runts on a genuinely half-duplex link is life as designed. Runts on a full-duplex port are a famous clue. The far end is aborting transmissions because it thinks it's colliding. That is the signature of a duplex mismatch, seen from the full-duplex side. The duplex guide milks this clue for everything it's worth.


Giants: frames too big to be legal

The opposite offense: frames larger than the port's configured maximum. Here's the twist — giants are usually a configuration problem, not a damaged-signal problem. One side is sending jumbo frames (oversized frames used to move bulk data more efficiently) or adding VLAN tags. The other side isn't configured to accept the larger size. True corruption occasionally manufactures a giant. But if this counter climbs steadily while CRC stays clean, go audit maximum-frame-size settings on both ends before blaming any cable.


Symbol errors: unreadable letters on the wire

Faster Ethernet doesn't put your raw bits on the wire directly. It first encodes them into symbols. Think of them as the letters of the alphabet the two ports agreed to write in. A symbol error means the hardware received a letter that isn't in the alphabet: the signal was distorted below even the decoder's tolerance. Symbol errors are wonderfully low-level — they fire on signal-quality problems sometimes even between frames. That makes them one of the earliest, most sensitive indicators of a degrading cable or optic. Not every platform exposes them; when yours does, treat them as the canary.


Carrier transitions: the link keeps blinking

Counted whenever the link itself goes down or comes up — some platforms call it carrier transitions, link flaps, or interface resets. Each pair of transitions is a moment the two ports lost each other entirely. The cause might be a loose or oxidized connector, a cable pinched to the edge of failure, a dying transceiver, unstable power at the far end, or somebody's foot. A handful at deployment time is normal (cables get plugged in). A count that keeps growing on an untouched link is a physical problem in progress. Each blink resets user sessions. So this counter is behind an enormous share of "I keep getting disconnected" tickets. It is also behind the maddening evidence-free failures covered in why intermittent problems are hard.


Collisions and late collisions: echoes of a shared past

On ancient shared Ethernet, all devices shared one channel, politely detecting when two talked at once — a collision — and retrying. Ordinary collisions on a half-duplex link were normal life. On today's switched, full-duplex links there is no shared channel. So any collision count above zero means someone thinks it's still the old days: almost always a duplex mismatch. The aggravated form is the late collision — a collision detected after the frame's first 64 bytes. The rules say this should be impossible on a properly built link. Late collisions mean the frame is not retried by the hardware; the data just dies. Their classic causes: a duplex mismatch (seen from the half-duplex side), or a cable run stretched far beyond its legal length.


One neighboring counter deserves a boundary sign. Discards are frames the device dropped on purpose — usually a congested queue, perfectly healthy frames with nowhere to go. Confusing them with errors sends you hunting cables when you should be hunting bandwidth. So read the guide to errors, discards, and drops before you alert on either.


The quick-reference table


CounterPhysical meaningPrime suspects
CRC / FCS errorsFrame arrived with flipped bitsNoise, bad cable, weak optic, duplex mismatch
Alignment errorsFrame mangled mid-byteCollisions, serious noise; travels with CRC
RuntsAborted frame stumps under 64 bytesCollisions; mismatch (full-duplex side)
GiantsFrames over the configured maximumJumbo/VLAN config mismatch, rarely damage
Symbol errorsUndecodable signal on the wireDegrading cable or optic; earliest warning
Carrier transitionsLink dropped and returnedLoose/failing connector, dying optic, far-end power
Collisions (any, on full duplex)Two talkers on a link built for two lanesDuplex mismatch
Late collisionsCollision after the legal window; frame lostDuplex mismatch (half-duplex side), overlong run

Totals lie: read deltas instead


Back to the big quirk. Because counters are lifetime totals, the number itself answers almost nothing — the change in the number answers almost everything. The manual version: note the value, wait a fixed interval, read it again, subtract. Here's what that looks like polling one uplink's error counter over SNMP, five minutes apart:


10:00:03  IF-MIB::ifInErrors.24 = Counter32: 143611
10:05:03  IF-MIB::ifInErrors.24 = Counter32: 143802

Delta: 191 new errors in five minutes. Now you know something: this fault is active, and you even have a rate to watch rise or fall as you test fixes. This is also exactly how monitoring systems think. A good one polls each counter on a schedule and stores the series. It alerts on the delta per interval — "more than N new errors in five minutes" — rather than on the meaningless lifetime total. Alerting on totals produces either eternal silence or one alert that never clears. Alerting on deltas produces a bell that rings while the crime is happening.


What threshold? On modern switched links the honest baseline is zero. A healthy full-duplex port can run for months without a single error. So there's no "normal background level" to tune around. A sensible starting rule: warn on any nonzero delta repeating across multiple intervals on an important port. And treat single-digit blips at link bring-up (cables being plugged in, devices rebooting) as noise. You'll tighten this with experience, but "roughly zero, always" is genuinely the standard to hold cables to.


Tip: resist the urge to clear counters on shared equipment, tempting as a clean zero feels. Counters are shared state — the colleague troubleshooting tomorrow loses the history, and your own "before" number vanishes from the record. Write down the value and the time instead; subtraction is cheap. If you do clear (some platforms only show certain details that way), announce it in the ticket.


Steady trickle or bursts: two different diseases


Once you're watching deltas over hours or days, the shape of the errors starts to diagnose the fault for you. Two shapes cover most cases.


Shape one: a steady rate that scales with traffic. A marginal medium damages each bit with roughly constant probability. That could be an attenuated cable, a weak laser, or a dirty connector letting through a degraded signal. Push more bits, lose more frames, in proportion. The tell: divide each interval's error delta by its traffic volume. If that ratio holds roughly steady from quiet hours to busy hours, the medium itself is marginal. The fix lives in hardware: test the cable, read the light levels, clean or replace.


Shape two: bursts clustered in time, indifferent to traffic. Errors that arrive in clumps — hundreds in one minute, then hours of silence — while traffic stays flat point to an event, not a medium. That could be electrical interference from a motor starting, an elevator run, or a light ballast failing. Or a connector is being physically disturbed (a rolling chair, a closet visit, a vibrating rack). The tell is correlation with the physical world, so note the burst timestamps and go find what the building does at those moments. Bursts are the harder hunt — the evidence sleeps between events — but the timestamps turn a vague gremlin into a stakeout with a schedule.


There's also a reassuring third shape worth naming: the frozen total. A large lifetime count whose delta stays at zero interval after interval is history, not disease. It might reflect a bad patch cord replaced long ago, one rough afternoon during construction, or a device that rebooted mid-transfer. Don't ignore it entirely (note it in documentation so the next admin doesn't re-panic), but don't chase it either. The delta discipline exists precisely so old scars stop triggering new investigations.


A worked diagnosis


Let's run one end to end. The complaint: "the third floor is slow, sometimes." The suspect path: the uplink from the floor's access switch to the core, port 24. You start polling ifInErrors on the core side every five minutes and let it run a day. Selected readings:


TimeError totalDelta (5 min)Traffic in
03:00151,004+24 Mb/s
09:00153,891+214420 Mb/s
09:05154,096+205410 Mb/s
13:00158,410+96190 Mb/s
19:00160,102+1122 Mb/s

Read it like the guide taught you. First, the fault is active — deltas are nonzero today, so this isn't ancient history. Second, the shape: errors track traffic beautifully. There is roughly one error per two megabits, whether it's the overnight trickle or the morning rush. That is the constant-ratio signature of a marginal medium, not the clustered-burst signature of interference. Third, direction: these are input errors on the core side, so frames are being damaged traveling access-to-core. The access switch's facing port shows clean input, confirming the hurt direction.


A marginal medium on a fiber uplink says: go read the light. The transceiver's diagnostics show receive power on the core side scraping along just above its alarm floor. That is a weak incoming signal, exactly what the counters predicted. The optics guide shows how to read those numbers. The closet log shows a re-patch on that path recently; a dirty connector is the odds-on cause. Clean the connectors, re-check: receive power jumps healthy, and the error deltas flatline at zero through the next morning's rush. Write the story in the ticket, keep the graph. Total hands-on time: minutes — because the counters did the localizing before you ever touched hardware.


Polling and alerting in practice


To turn all this into standing infrastructure rather than a heroic one-off, work through this checklist:


  • Choose the ports that matter first: uplinks between switches, ports feeding servers and storage, firewall and router links. Every port on every switch can come later.
  • Poll the right objects: in and out errors and the link state for every monitored port. Also poll the detailed Ethernet counters (CRC, collisions, symbol errors) where the platform exposes them over SNMP.
  • Pick a sane interval: five minutes is the customary sweet spot — fine-grained enough to catch and time bursts, gentle enough on the devices.
  • Alert on deltas, not totals: warn when new errors appear across consecutive intervals; alert on any link flap of an uplink; stay quiet about lifetime totals.
  • Keep the history: weeks of retained counter series are what make the steady-versus-burst diagnosis possible at all. They're your before-and-after proof when you fix something.
  • Watch after every change: re-patches, new cables, new transceivers. The first day of counters after maintenance is the cheapest acceptance test you will ever run.

If you're wondering whether all this counter-reading might be overkill: it's the opposite. It is the cheapest diagnostic layer you own — already collected, already per-port, already historical the moment you start polling. Most of the expensive mysteries in the symptoms guide dissolve into ten minutes of counter arithmetic.


Where to go next


You can now read the diary: what each counter physically means, and why deltas beat totals. You also know how the shape of errors over time separates a fading cable from a haunted elevator shaft. Two counters kept pointing at one particular villain — runts on one side, late collisions on the other. That villain gets the full treatment in the duplex mismatch guide. For the fiber-flavored version of "marginal medium," read monitoring optics. For what to do with your hands once counters implicate copper, see cable plant troubleshooting. The whole series lives at Finding Physical-Layer Problems.


Put this into practice with SysPeep Monitor

Everything in this guide comes down to polling counters on a schedule and alerting on the deltas. That is exactly the boring, reliable work monitoring software exists to do. SysPeep Monitor discovers your switches. It polls interface error counters and link state over SNMP. It keeps the history and alerts when the numbers start moving. The free trial runs on any Windows machine.

Download Free TrialLearn About SysPeep