Fixing Loss: The Remediation Catalog
Diagnosis without treatment is just expensive sightseeing. By now you have confirmed the loss is real, found roughly where it lives, and read the fingerprints it left in the counters. This guide is the repair manual. It covers the fixes that actually cure packet loss on small and mid-size networks. They are ordered by a blunt, practical rule — cheapest and most likely first. You will replace far more patch cables than switches, and fix duplex mismatches far more often than you will win arguments with an ISP.
Every entry follows the same shape. First comes when this is the right fix (the fingerprint that points here), then the steps. Then comes the part most people skip: the verification test that proves it worked. A fix without verification is a superstition with a receipt. If the fingerprints are unfamiliar, the guide to errors, discards, and drops is the decoder ring for everything below.
In this guide
- First, the discipline: measure, change one thing, measure again
- Fix 1: Cables, connectors, and optics
- Fix 2: Duplex and speed mismatches
- Fix 3: Congestion relief: QoS, scheduling, rate limits, upgrades
- Fix 4: Buffer tuning — helpful, with cautions
- Fix 5: Wi-Fi remedies
- Fixes 6 and 7: Drivers, firmware, and dying hardware
- Fix 8: Escalating to your ISP with evidence
- The printable fix order
First, the discipline: measure, change one thing, measure again
Before the catalog, the ritual that makes every entry work: four steps, no shortcuts. One: take a baseline. Measure the loss with the same test you will use afterward, whether that is a long ping, a counter snapshot, or both. The guide to measuring loss accurately covers how to do this without fooling yourself. Two: make exactly one change. Three: repeat the identical measurement, and run it long enough to cover the problem's schedule. A fault that appears at 8:40 on weekday mornings cannot be declared cured by a clean Saturday test. The article on the intermittent-loss workflow exists for exactly this. Four: write it down: date, time, what changed, numbers before, numbers after.
The two-change trap: swap the cable and update the driver in one visit, and when the problem stops you will never know which one fixed it. So when it reappears on another port next quarter, you get to do the whole investigation again. One change, one measurement. It feels slower. It is faster.
With the ritual established, match your evidence to a starting point:
| What the evidence shows | Start with |
|---|---|
| Input/CRC errors climbing on one link | Fix 1: cables, connectors, optics |
| Late collisions; loss that grows with load on one link | Fix 2: duplex and speed |
| Output discards on a busy egress at peak times | Fix 3: congestion relief |
| Discards despite low average utilization (microbursts) | Fix 4: buffers |
| Only wireless users affected | Fix 5: Wi-Fi remedies |
| One host suffers; its switch port counters are clean | Fix 6: drivers and firmware |
| Errors on many ports at once; unexplained reboots; heat correlation | Fix 7: hardware replacement |
| Loss begins past your WAN handoff; your side is clean | Fix 8: ISP escalation |
Fix 1: Cables, connectors, and optics
When it is the right fix: input errors or CRC errors (frames arriving damaged) climbing on one specific interface, loss confined to one link, sometimes worse when the closet runs warm. The most common cause of loss on a LAN and the cheapest to fix — always suspect the physical layer before anything clever.
The steps, in escalating order of cost:
- Reseat both ends. Unplug, inspect the clip and pins, plug back firmly. Free, and it resolves more tickets than anyone likes to admit.
- Swap the patch cable for a known-good one — a fresh cable, not another veteran from the drawer of mystery cables.
- Inspect the run: look for kinks and crush damage (chair wheels and door thresholds are serial offenders). Check for staples through the jacket, tight zip ties, and long runs parallel to power cables. Also check for total length pushing past the roughly 100 m copper limit.
- On fiber: swap the transceiver (the small plug-in optic module, often called an SFP). Clean the connector ends with proper fiber tools — a speck of dust is a wall at these wavelengths. Try the spare strand pair if one exists.
- Try a different switch port. Ports fail too; if the errors follow the port rather than the cable, you have skipped ahead to Fix 7.
Verification: note the error counters, or clear them — either works — then pass real traffic or simply wait a day:
# typical switch syntax (varies by vendor) show interfaces counters errors clear counters gigabitethernet0/10 # on a Linux host at the other end of the link ethtool -S eth0 | grep -i -e err -e drop
Suppose a link was producing hundreds of errors an hour and now shows a delta of zero for a full day. If that includes the hours when the problem used to appear, it is fixed. Errors merely reduced means you improved a symptom without curing the fault; keep going.
Fix 2: Duplex and speed mismatches
When it is the right fix: the giveaway fingerprint is loss that scales with traffic on one link. Quiet links test clean; busy links fall apart. There are late collisions counting up on one side and CRC errors on the other. Duplex is whether a link talks both directions at once (full) or takes turns like a walkie-talkie (half). A mismatch — one end full, one end half — happens when one side is hard-coded and the other is left to autonegotiate. Hearing no negotiation partner, the auto side falls back to half duplex by standard behavior. The ends now disagree about the rules of the road, and every simultaneous transmission mangles frames.
The steps. First, look at both ends of the suspect link — both, always, because the mismatch is the disagreement:
# on the switch (typical syntax)
show interfaces status
Port Name Status Duplex Speed
Gi0/10 uplink connected a-full a-1000 ("a-" = autonegotiated)
# on a Windows host
Get-NetAdapter | Format-Table Name, Status, LinkSpeed, FullDuplex
# on a Linux host
ethtool eth0
Speed: 1000Mb/s
Duplex: Full
Auto-negotiation: on
Then the modern rule: set both ends to autonegotiate. On current equipment it is reliable, the default, and at gigabit and faster effectively required. The old folklore about hard-coding "for stability" dates from buggy early implementations; today it mostly creates the problem it was meant to prevent. If a stubborn museum piece will not negotiate, force both ends identically, label the port loudly, and plan its retirement. What you must never do is force one side and leave the other on auto — that is the exact recipe for the fault.
Verification: after bouncing the link, confirm both ends report full duplex at the expected speed. Then run the load test a mismatch fails: a sustained large file copy across the link while pinging through it. Zero loss during the transfer, and late-collision counters frozen, means cured.
Fix 3: Congestion relief: QoS, scheduling, rate limits, upgrades
When it is the right fix: output discards climb on the egress of a bottleneck during busy windows. Typically that is the WAN uplink or a switch uplink. Utilization graphs touch the ceiling at the same times. Nothing is broken; the pipe is full, the queue overflows, and the overflow is dropped. Four levers, usually pulled in this order.
3a. Give real-time traffic priority (QoS)
QoS (quality of service) is the set of rules deciding who waits and who is dropped when a link is full. The standard move: put voice and video into a priority queue at the bottleneck. That way, a briefly full link starves the file transfer instead of the phone call. Most phones and conferencing systems already mark their packets (a field called DSCP). Your job is to make the bottleneck device honor the marks. Two honest caveats. QoS adds no capacity — it only chooses victims wisely. If the link is saturated all day, you need lever 3d. And it only works where you control the queue: your router's upstream egress, yes; the ISP's downstream toward you, only if they cooperate. Verify: during the worst window, priority traffic tests clean while bulk transfers slow down — the system working as designed.
3b. Move the bulk elephants
Backups, offsite replication, patch distribution, archive syncs — schedule them into genuinely idle hours, and alert on overruns. Datasets grow, and a 1:00 AM job will eventually still be running at 8:40. Verify: discard deltas at zero across the business day for a full week.
3c. Rate-limit the hog
When one system — a cloud sync client, a camera archiver, someone's video habit — owns the bottleneck, cap it below the danger line at the nearest device that supports limits. Per-port counters identify the hog roughly; flow analysis (covered elsewhere in the topics library) identifies it precisely. Verify: the hog's port plateaus at its cap; discards on the bottleneck stop.
3d. Upgrade the link
Sometimes the honest diagnosis is that demand has outgrown the pipe: utilization high most working hours, discards daily, no single villain to cap. Then the fix is capacity. The bandwidth guides in the library cover how to size it from your utilization data rather than from guesswork. Verify: same as ever — the busy-hour measurements, repeated on the new link, come back clean.
Fix 4: Buffer tuning — helpful, with cautions
When it is the right fix: a buffer is the waiting room at each port's exit; when it fills, arrivals are dropped and the output-discard counter ticks. The fingerprint that points here is discards on a link whose average utilization looks innocent. That is the signature of microbursts, split-second pileups (a dozen sources answering one client at once). They overflow a small buffer and vanish before any graph can see them. Modestly larger queues on that port help absorb those bursts.
Now the cautions, because this fix has a dark side. Where congestion is persistent rather than momentary, a bigger buffer does not prevent the drop. It delays it, and meanwhile every packet waits in a longer line. Taken to extremes this is bufferbloat: loss transformed into seconds of latency, which for real-time traffic is a worse disease than the loss was. A voice packet is better dropped promptly than delivered late; TCP, for its part, relies on timely drops as its signal to slow down. Deep buffers can quietly remove the feedback that keeps senders honest.
The steps: treat buffers as the last congestion lever, after Fix 3 has done what it can. Record current values before touching anything, change one queue on one port, and consult your vendor's documentation. The knobs differ wildly between platforms. Verification is double here: discards must go down and latency under load must not balloon. Measure both: run a ping through the link while a heavy transfer saturates it, before and after. Fewer drops with round-trip times still sane is a win. Fewer drops with latency ten times worse means you have hidden the problem, not fixed it. In that case, put the setting back.
Fix 5: Wi-Fi remedies
When it is the right fix: only wireless users suffer. The paired wired-versus-wireless test from the Wi-Fi loss guide settles that. The access points also report high retry rates or busy channels. That guide explains the why; here is the what-to-do:
- Fix the channel plan. Put neighboring APs on non-overlapping channels. In the 2.4 GHz band that means 1, 6, and 11 in most regions. Use sensible channel widths rather than the widest available. Overlap between your own APs is the most common self-inflicted wireless wound.
- Reconsider power and placement. Maximum transmit power is not a gift: it makes every AP's "room" overlap its neighbors', increasing contention. Moderate power with even placement beats a few APs shouting. Move APs away from metal, thick masonry, and the microwave's line of fire; put them where the people are.
- Retire the lowest legacy rates. Disabling the slowest legacy data rates on the APs stops distant or ancient clients from monologuing in slow motion. It forces earlier roaming to a closer AP. Caution: inventory first — the venerable barcode scanner or lab instrument may support nothing newer.
- Split the crowd. Where one AP carries a meeting room's worth of devices, add coverage — on a clean channel plan — rather than turning up the volume.
Verification: at the same busy hour that used to hurt, retry rates and channel utilization drop on the AP dashboard. At that hour, the paired ping test also shows the wireless laptop tracking close to its wired control.
Fixes 6 and 7: Drivers, firmware, and dying hardware
Fix 6: Update NIC drivers and firmware on the suspect device
When it is the right fix: one host suffers loss while its neighbors on the same switch are fine. And — the clincher — the switch port's counters for that host are clean. Packets are dying inside the machine: a buggy network card driver, broken offload features, or stale firmware. Classic tells: it started after an operating system update, or the machine is brand new (shipping drivers are often the oldest thing in the box).
The steps: install the network adapter driver from the NIC or system vendor rather than the operating system's bundled generic. Apply NIC firmware if offered. If trouble persists, temporarily toggle the adapter's offload features (checksum and large-send offload) as a diagnostic. These settings push work onto the NIC and occasionally misfire. Revert whatever changes nothing. Last resort on this branch: a different NIC, port, or dock. Verify: a long timestamped ping log from that host plus a sustained transfer, clean across a full day.
Fix 7: Replace failing hardware
When it is the right fix: the evidence stops pointing at any one link and starts pointing at the box. Signs include errors climbing on many ports at once, ports that recover after a reboot and sicken again, and symptoms tracking closet temperature. Other signs include fan or power-supply warnings in the logs, memory complaints, or uptime resets nobody scheduled. Switches rarely die dramatically; they get strange first.
The steps: first the cheap test — move a critical link to a spare port and see if its troubles follow the port or stay with the cable. Then swap in a spare or replacement unit. Keep the suspect device labeled and quarantined for warranty, not back in the spares drawer to ambush a future you. Verify: the replacement carries identical traffic for a week with zero error deltas.
Fix 8: Escalating to your ISP with evidence
When it is the right fix: your path bisection shows loss beginning past your WAN handoff. The path is clean to your router's inside, lossy from the first ISP hop outward. Your own WAN interface counters are also spotless. Now the fix is persuasion, and the currency of persuasion is evidence that cannot be waved away. Gather, before you call:
- Timestamped ping logs to the ISP's gateway spanning several days, covering both bad windows and clean ones.
- The same logs to a second public target, proving it is the path and not one destination having a bad week.
- Per-hop reports from bad windows and good ones, so the difference is visible.
- Your WAN interface counters showing zero errors and discards on your side.
- Your circuit ID, account number, and exact timestamps with your time zone.
# 600 probe cycles (about 10 minutes), summarized per hop mtr -rwb -c 600 203.0.113.1 > mtr-tuesday-0830.txt # Windows built-in alternative (slower, same idea) pathping 203.0.113.1
Then present it the way you would want to receive it: summary up front, data attached, one specific request. Something like:
Circuit FIBER-4471, account 0182337 Symptom: 3-6% packet loss to your gateway (203.0.113.1), weekday mornings 08:00-10:00 local time; clean otherwise. Evidence attached: - ping-gw.csv: one ping per 2 seconds, 9 days, timestamped - mtr reports from three bad mornings and three clean afternoons - our WAN interface counters: zero errors/discards our side Already done: router rebooted, cables reseated, tested from a wired host connected directly to the router. No change. Request: a line test on this circuit during a morning window, and a ticket number for this report.
The "already done" line does quiet heavy lifting: it answers the first-line script — reboot the router, check your cables — before it is read aloud. Expect to repeat yourself; reference the ticket number and resend the same data, updated. Calm, specific, relentless is the winning register. And be fair about the boundary: loss deep in the wider internet is outside their reach. But the access circuit and their gateway are theirs, and that is usually where your evidence points.
Tip: run every piece of ISP evidence from a wired machine, ideally one plugged directly into your router. One test taken over Wi-Fi hands them an honest reason to dismiss the entire pack. Wireless adds its own loss and jitter, and their engineers know it.
Verification: after they report a fix — a line rebalanced, an optic replaced, a port moved — run the identical logs for another week. The before-and-after files are also your record if the fault returns next quarter.
The printable fix order
The whole catalog on one card — pin it inside the server closet door:
| Order | Fix | Reach for it when | Proof it worked |
|---|---|---|---|
| 1 | Reseat / replace cable, swap optic | Input/CRC errors on one link | Error delta zero for a full day |
| 2 | Correct duplex and speed (auto both ends) | Late collisions; loss grows with load | Clean ping during a sustained transfer |
| 3 | QoS, reschedule bulk, rate-limit, upgrade | Output discards at peak hours | No discards, clean ping at the worst hour |
| 4 | Buffer tuning (cautiously) | Discards despite low average utilization | Discards down and latency under load flat |
| 5 | Wi-Fi: channels, placement, legacy rates | Wireless-only symptoms, high retries | Retry % and utilization drop at busy hour |
| 6 | NIC drivers and firmware | One host; its switch port counters clean | Day-long ping log from that host clean |
| 7 | Replace failing hardware | Errors on many ports, reboots, heat | A clean week on the replacement |
| 8 | ISP escalation with evidence | Loss starts past your handoff | Same logs clean for a week after their fix |
And beneath the card, the ritual: baseline, one change, the same measurement again, notes. Loss you can measure is loss you can fix. And loss measured before and after is loss you can prove you fixed. That is what turns a lucky afternoon into a reliable network. For the rest of the story — where loss comes from, how much matters, and how to catch the kind that hides — the Packet Loss pillar links every guide in this series.
Put this into practice with SysPeep Monitor
Half of fixing is proving the fix, and that takes before-and-after data. SysPeep Monitor collects it automatically: interface error and discard trends and ping response history. It also alerts if the numbers ever creep back up after your one careful change. Make the fix, then watch the graphs go flat. The free trial installs on a Windows machine in minutes.
Download Free TrialLearn About SysPeep