Home › Topics › Packet Loss › Interface Counters

Errors, Discards, Drops: Reading Interface Counters Correctly


Every port on every managed switch, router, and firewall you own carries a little flight recorder. It has been running since the device booted. It costs nothing. And it answers the question every packet-loss hunt eventually asks: did this port lose traffic, and if so, how? These are the interface counters — errors and discards, input and output. And the strange thing about them is how rarely anyone reads them. Admins will reboot switches, replace cables, and open ISP tickets while the exact evidence sits uninspected on the very port in question.


Part of the reason is that the counters look interchangeable. "Errors," "discards," "drops" — surely all just flavors of bad? They are not. Each of the four counter families records a different way for a frame to die. The pattern of which counters climb — and which stay at zero — distinguishes a damaged cable from a congested uplink from a misconfigured duplex setting. Often, it does so in under a minute, from your desk.


This is the reference guide for those four families. You will learn what each counter genuinely means and which combinations point to which faults. You will learn how to read the counters at the CLI and remotely over SNMP. And you will learn the two habits — thinking in growth rates, and baselining old dirt — that separate people who read counters correctly from people who panic at big numbers.


In this guide

  • Where the counters live, and a first annotated look
  • The four counter families, precisely
  • Combinations: the diagnostic grammar
  • Reading counters over SNMP
  • Rate versus absolute: growth is the signal
  • Old dirt: baselining and clearing
  • The triage table
  • Where to go next

Where the counters live, and a first annotated look


The counters are kept by each device, per port, updated in hardware as frames flow. You can read them three ways. You can look on the device's command line, on a host's operating system (for its own NICs), or remotely over SNMP. That last method is how monitoring systems collect them. Here is what a typical switch CLI shows for one port. The exact wording and layout vary by vendor, but these fields exist almost everywhere:


GigabitEthernet0/24 is up, line protocol is up
  Description: uplink-to-core
  Full-duplex, 1000Mb/s
  ...
  184023455 packets input, 213 input errors      <-- family 1: arrived damaged
     198 CRC, 15 frame, 0 runts, 0 giants        <-- what kind of damage
  412 input discards                             <-- family 2: arrived fine,
                                                     thrown away anyway
  391208871 packets output, 0 output errors      <-- family 3: failed to send
     0 collisions, 0 late collisions
  Total output drops: 48211                      <-- family 4: queue full
                                                     on the way out

Four families, four different deaths. Notice also the two plain packet totals — 184 million in, 391 million out. They are the denominators that give every other number meaning. A count of 213 damaged frames against 184 million received is a rounding error. A count of 48,211 drops against 391 million sent is worth a look at when they happened. Keep both halves in view whenever you read a port.


Servers and workstations keep the same books for their own network cards. On Linux, ip -s link show eth0 prints per-direction errors and drops; on Windows, netstat -e summarizes the same families:


C:\> netstat -e
Interface Statistics
                           Received            Sent
Bytes                    8123456789      5432109876
Unicast packets             9876543         7654321
Discards                        448               0
Errors                          213               0

Reading both ends of a link — the switch's view and the host's — is worth the extra minute. Each end only counts what it receives, so the two views together cover both directions of the cable. And a disagreement between them is itself a clue about which direction is suffering.


The four counter families, precisely


Family 1: input errors — frames that arrived damaged

An input error means a frame reached this port physically broken. The most common subtype is the CRC error. Every Ethernet frame ends with a checksum computed by the sender. If the receiver's recalculation disagrees, at least one bit changed in transit. In that case, the frame cannot be trusted and is destroyed on arrival. Its siblings describe other malformations. With alignment/frame errors, the frame did not end on a clean byte boundary. The runts are impossibly short fragments, classically collision debris. And giants are frames larger than the port's configured maximum. Those usually mean an MTU or VLAN-tagging mismatch rather than damage.


The mental model: input errors mean the physical layer is lying. The damage happened somewhere between the far end's transmitter and this port's receiver. Possible sources include a cable, a connector, a patch panel, an optic, or electrical interference. Others include a failing NIC on the far side or the corrupted-fragment side of a duplex mismatch. The port that counts the errors is the witness, not the culprit: it is faithfully reporting the noise it hears from upstream.


Family 2: input discards — arrived intact, thrown away anyway

An input discard is stranger: the frame arrived perfect — checksum valid, format clean — and the device threw it away regardless. The standard reason is resource pressure. The inbound queue or buffer was full because the device could not process arrivals fast enough, so something had to go. On various platforms, certain policy decisions can also land in this bucket. Examples include a frame for a VLAN the port does not carry, or traffic an ACL rejects. Vendors differ on exactly what they count here, so treat the precise membership as device-specific.


The mental model: input discards mean the wire is fine and the device is the bottleneck (or the policy). Rising input discards with clean error counters is never a cabling problem. It is the box itself struggling to keep up, or dropping what it was told to drop.


Family 3: output errors — the port failed to transmit

An output error means the port tried to send a frame and could not. On modern full-duplex links this should essentially never happen. That makes the counter valuable precisely because it is usually zero. When it climbs, suspect hardware — a failing port, a dying transceiver. Or suspect the one classic protocol-level cause that refuses to die: duplex mismatch. On the half-duplex side of a mismatched link, transmissions keep colliding with the far end's traffic. The giveaway subtype is the late collision. This is a collision detected after the frame's first 64 bytes, which correctly configured links do not produce.


The mental model: output errors are rare, and rare is informative. A port with growing output errors has a hardware problem or a duplex problem, and the late-collision sub-counter votes strongly for duplex.


Family 4: output discards — the queue was full on the way out

An output discard (many CLIs say "output drops") means the frame was healthy and the device knew exactly where to send it. But the output port's queue was full, so the frame was dropped on the way out the door. This is the signature of congestion. More traffic wanted to leave through that port than the port could carry. The buffer absorbed what it could, and the overflow died. Oversubscribed uplinks write their confession in this counter. So do fast-to-slow speed steps (a ten-gigabit server feeding a one-gigabit client port). So do millisecond microbursts that never show on minute-averaged graphs.


The mental model: output discards mean too much traffic for the exit, full stop. Nothing is broken, which is exactly why replacing hardware never fixes it. This is the most commonly misread counter of the four — and, on busy networks, the most commonly nonzero.


Combinations: the diagnostic grammar


Single counters are words; combinations are sentences. These are the patterns worth memorizing:


  • Input errors on one port, everything else clean: that specific link's physical path — patch cable, wall jack, patch panel, optic, or the far device's NIC. Cheap fixes first: reseat, clean, swap the cable, then swap the optic.
  • Input errors on many ports of one switch: stop blaming individual cables. Think shared factors — electrical interference in that closet, a common cable bundle pinched during last week's tidy-up, or the switch itself beginning to fail.
  • Output discards on an uplink: congestion. Correlate with utilization and time of day; the cure is capacity, traffic engineering, or QoS — never a cable cart.
  • Discards (either direction) with zero errors anywhere: a capacity or policy story by definition — the frames were perfect. Investigate load, buffers, and configured drop policies.
  • Output errors with late collisions: duplex mismatch until proven otherwise. Read speed and duplex on both ends of the link; expect to find one side hard-coded.
  • Errors and discards climbing together: possibly two independent problems — but note that an error-riddled link also triggers TCP retransmissions, inflating load elsewhere. Fix the errors first; then re-measure before judging the discards.

Notice what the grammar gives you: errors point at the physical world, discards point at capacity and policy. That one sentence, applied before any hardware is touched, prevents most wasted afternoons. (For the full cause-by-cause tour behind these fingerprints, see where packets actually die.)


Reading counters over SNMP


The CLI is fine for one port on one afternoon. For continuous visibility you read the same counters remotely over SNMP, which is precisely what monitoring systems do. The four families live in the standard interfaces table (IF-MIB) that virtually every managed device implements, one value per port:


Counter familySNMP nameNumeric OID (per port)
Input errorsifInErrors1.3.6.1.2.1.2.2.1.14
Input discardsifInDiscards1.3.6.1.2.1.2.2.1.13
Output errorsifOutErrors1.3.6.1.2.1.2.2.1.20
Output discardsifOutDiscards1.3.6.1.2.1.2.2.1.19

Each OID gains the port's index number on the end, and a walk returns the whole switch at once:


snmpwalk -v2c -c labread 192.168.1.2 IF-MIB::ifOutDiscards

IF-MIB::ifOutDiscards.1  = Counter32: 0
IF-MIB::ifOutDiscards.2  = Counter32: 0
IF-MIB::ifOutDiscards.25 = Counter32: 48211   <-- port index 25: the uplink,
...                                               telling the congestion story

Walk IF-MIB::ifDescr once to map index numbers to port names, and the picture completes. Finer-grained physical detail (the CRC-versus-alignment split) lives in the Ethernet-specific MIB as dot3StatsFCSErrors and dot3StatsAlignmentErrors on devices that implement it. But the four IF-MIB counters above carry the bulk of the diagnostic weight. They are what your monitoring should be polling on every port that matters. A sensible polling interval is every one to five minutes. That is frequent enough to timestamp when growth begins, and gentle enough that even a modest switch will not notice the questions.


Rate versus absolute: growth is the signal


Now the habit that changes everything about how you read these numbers: every counter is a lifetime total, accumulating since the device booted (or the counter was cleared). A port showing 1,834,220 input errors sounds catastrophic. But if the switch has been up for three years and those errors date from a cable fixed two winters ago, the port is perfectly healthy today. Meanwhile a port showing a humble 9,000 errors that were all counted this morning is on fire. The absolute value answers "has this port ever suffered?" The useful question is "is it suffering now?" — and that is a question about growth.


Measuring growth takes two readings and a subtraction. Read the counter, note the time, come back later, read again:


09:00  ifInErrors.12 = 1834220
10:00  ifInErrors.12 = 1834220     <-- +0 per hour: historic scar, healthy today

09:00  ifInErrors.7  = 1204
10:00  ifInErrors.7  = 9873        <-- +8669 per hour: active problem, act now

What growth rate is acceptable? On a healthy modern link the honest target for errors is zero per hour. A trickle of new CRC errors on a busy port means a marginal physical path that will get worse. For discards, judge growth against traffic and timing. Discards that grow only during the nightly backup are a known squeeze. Discards growing all day on a core uplink are a capacity project announcing itself. And always keep the denominator nearby. A count of 8,669 new errors against 30 million packets in the same hour is a 0.03% damage rate. That is small in ratio but still a defect worth fixing on a LAN, where the standard is clean. (What loss rates mean for actual users is the subject of the acceptable-loss guide.)


Tip: this is the single strongest argument for monitoring these counters instead of visiting them. A poller that reads every port every few minutes turns lifetime totals into growth-rate graphs automatically. A graph that has been flat for months makes this morning's uptick unmissable — and timestamps it for you.


Old dirt: baselining and clearing


Because counters are lifetime totals, mature networks accumulate historic dirt. Some ports' big scary numbers describe problems long since fixed — or never real. An example is the error burst every unplug-and-replug writes while the link renegotiates. Two habits keep the history from fooling you (and your colleagues) later.


First, reset or record after every fix. When you replace a cable or correct a duplex setting, clear the port's counters. Most switch CLIs have a per-interface clear command. Or — just as good — write down the counter value and the time after the fix. Tomorrow's question, "has it grown since the fix?", becomes answerable at a glance. If you clear, note one honest wrinkle. On many devices the CLI clear resets only the displayed totals, while the raw SNMP counters march on untouched. So your monitoring may disagree with the CLI. Growth-based monitoring shrugs this off, which is one more point in its favor.


Second, baseline the dirty ports. Keep a note — even a text file — of which ports carry known historic counts and why ("Gi0/7: ~1.8M input errors, patch cable replaced, counter not cleared"). It inoculates the whole team against rediscovering the same scar during every future incident. And before celebrating any port that reads zero, glance at the device's uptime. Counters also reset to zero on every reboot. A switch that restarted last night has an alibi, not a clean record.


A sixty-second port health read


Everything above compresses into a routine you can run on any suspect link faster than a switch can reboot:


  • Step 1: identify the exact ports the suffering traffic crosses — and remember every link has two ends worth reading.
  • Step 2: read all four counter families on each port, plus the packet totals (your denominators) and the device's uptime (your time frame).
  • Step 3: apply the grammar: errors point physical, discards point at capacity or policy, late collisions shout duplex.
  • Step 4: establish growth — re-read after a coffee, or better, consult the monitoring graph that has been watching all along.
  • Step 5: write down what you saw, with values and times, so the next reading — yours or a colleague's — starts from evidence instead of folklore.

The triage table


Working direction matters: users bring you symptoms, not counters. This table starts from the complaint and tells you where to look:


SymptomCounters to checkLikely cause
One wired user slow or flakyInput errors on their access port (both ends of the cable)Damaged patch cable, jack, or NIC; duplex if late collisions appear
Everyone slow at busy timesOutput discards on uplinks, checked against utilizationCongestion / oversubscription
Calls between sites choppyOutput discards on the WAN-facing egress portUndersized circuit or missing QoS on the squeeze point
Loss on a fiber runInput errors (CRC) at both endsDying optic or dirty connector — clean, reseat, then swap
Server "up but slow" after maintenanceOutput errors + late collisions; duplex readings both endsDuplex mismatch introduced by a hard-coded setting
Loss reported, all error counters zeroDiscards on every hop along the pathCapacity or QoS policy — not cabling; do not replace hardware
Errors on many ports, one closetInput errors across the switch; device healthInterference or power issue in the closet, a disturbed cable bundle, or a failing switch

Where to go next


Four families, one grammar: input errors — the wire is lying; input discards — the device is the bottleneck; output errors — hardware or duplex; output discards — too much traffic for the exit. Add the two habits. Judge growth, not totals. And keep the old dirt catalogued. With those habits, you can extract more truth from sixty seconds of counter reading than from an afternoon of speculative rebooting. From here, the causes tour connects each fingerprint to its full story. And the measurement guide shows where counters fit among the other loss-hunting tools. And the remediation catalog covers what to do about each confirmed cause. The complete set lives in the Packet Loss topic.


Put this into practice with SysPeep Monitor

These counters are exactly what SysPeep Monitor watches all day. It auto-discovers your devices and interfaces and polls errors and discards over SNMP. It turns lifetime totals into growth rates and alerts you when a flat line starts climbing. It has the history to show precisely when it began. The free trial installs on any Windows machine in minutes.

Download Free TrialLearn About SysPeep