Home › Topics › Packet Loss › Acceptable Loss

How Much Packet Loss Is Acceptable? Thresholds by Traffic Type


Sooner or later somebody — a boss, a vendor, an alert-configuration screen — will ask you how much packet loss is acceptable. You will discover that the two most popular answers are both wrong. The first wrong answer is "zero," usually delivered with great confidence. The second is "ah, a little loss is fine," usually delivered while a sales call drops for the third time.


The honest answer is that acceptability depends on three things. These are what kind of traffic is crossing the path, how long the path is, and over what time window you measured. A loss rate that a file download shrugs off can wreck a phone call. A loss rate that is harmless on a LAN can strangle throughput across a continent. And "1% loss" can mean a mild nuisance or a nightly outage depending on how the 1% is distributed.


This guide teaches you the reasoning, then gives you defensible working numbers for wired LANs, Wi-Fi, WAN and internet paths, and voice specifically. It finishes by turning those numbers into alert rules for a real small office. By the end, when someone asks the question, you will have an answer you can explain, which is worth far more than a number you memorized.


In this guide

  • Why "zero" and "a little is fine" are both wrong
  • How TCP hides loss — and what the hiding costs
  • Why real-time traffic feels loss instantly
  • Working thresholds by path type
  • The measurement window problem
  • When to alert: sustained versus momentary
  • A worked example: thresholds for a small office
  • Where to go next

Why "zero" and "a little is fine" are both wrong


"Zero" fails because IP networks are best effort by design. Every router and switch on a path is permitted to discard packets when its buffers fill. On any path you share with other people's traffic — which is every internet path — momentary overflow is a fact of life. Demand zero loss on a WAN link and you have signed yourself up to investigate ghosts forever. The occasional dropped packet during a traffic burst is the network working as designed, not a fault. (On your own switched LAN, "effectively zero" genuinely is the standard — more on that below — but as a universal target, zero is a fantasy.)


"A little is fine" fails because it ignores what the traffic is. The applications on your network split into two families that experience loss in completely different ways. One family — anything built on TCP: file transfers, web pages, email, most business apps — detects loss and re-sends the missing data. So users never see a gap; they see slowness. The other family — real-time traffic built on UDP: voice calls, video meetings, some game and streaming traffic — sends each packet exactly once. So every lost packet is a small hole in the experience that nothing ever fills. The same "little" loss is invisible to the first family and audible to the second. To set thresholds you can defend, you need to understand both mechanisms — they take one section each.


How TCP hides loss — and what the hiding costs


TCP (Transmission Control Protocol) treats the network as unreliable on principle. The receiver continuously acknowledges what has arrived. When the sender notices something missing, it re-sends. It may notice because acknowledgments arrive repeatedly asking for the same missing piece, or because an acknowledgment never comes at all. Your download completes byte-perfect over a path that quietly dropped hundreds of packets along the way. It is a genuinely beautiful piece of engineering, and it is also why loss problems fester: TCP hides the evidence.


The hiding has two costs. The first is latency: re-sent data takes at least one extra round trip to arrive. And the unlucky cases that need a full retransmission timeout stall the connection for far longer. Those occasional one-second freezes in an otherwise fine session are often exactly this. The second cost is bigger: TCP interprets loss as a congestion signal. The rules say that when packets are being dropped, the polite response is to slow down — sharply — and then speed back up only gradually. Lose packets steadily and TCP spends its whole life braking and cautiously re-accelerating, never reaching the speed the link could carry.


How bad is that in numbers? A classic back-of-envelope model describes steady TCP throughput. It says the achievable rate is roughly proportional to one over the round-trip time, times one over the square root of the loss rate. You do not need the formula's details — two consequences matter:


  • The same loss hurts long paths far more than short ones. The model's ceiling shrinks in direct proportion to round-trip time. A path with eighty times the RTT has, at the same loss rate, roughly one-eightieth the throughput ceiling. That is why a fraction of a percent of loss is a shrug on a 1 ms LAN and a strangled crawl on an 80 ms intercontinental link.
  • Improvements pay off on a square root. Cut the loss rate to a quarter and the ceiling only doubles. This is why "we reduced the loss a bit" so often produces disappointingly little visible improvement. It is also why the real goal on TCP-heavy paths is getting loss near zero, not merely lower.

Here is a feel for the scale of the first point, using nothing but the model's ratios. At the same loss rate, a path with a 1 ms round trip has a throughput ceiling about 20 times higher than a 20 ms path. The first path's ceiling is about 80 times higher than an 80 ms path. Loss percentage alone tells you almost nothing about TCP impact until you know the RTT next to it. (Round-trip time is a whole subject of its own — the Network Latency guides cover it properly.)


Why real-time traffic feels loss instantly


Now the other family. A voice call chops your speech into tiny packets, each carrying a small slice of sound. This is commonly about 20 milliseconds of audio per packet, fifty packets a second. These travel over UDP (User Datagram Protocol), which is TCP's minimalist sibling: no acknowledgments, no retransmission, no apology. This is not laziness — it is the correct design. If a packet holding 20 ms of Tuesday's sentence goes missing, re-sending it would deliver that sound after the moment it belonged to has already been played. Real-time traffic cannot use late data, so it does not bother asking for it.


Instead, a lost voice packet is simply a 20 ms hole in the audio. Phone systems paper over isolated holes with packet loss concealment. The receiver guesses at the missing slice, roughly by stretching the sound around it. One concealed packet here and there is genuinely inaudible. But concealment is a bandage, not a cure. Lose two or three packets in a row and the guesswork fails audibly. Words develop robotic warbles, syllables vanish, and the call turns into "you're breaking up." This is why the distribution of loss matters as much as the rate. 1% loss scattered evenly through a call may pass unnoticed. But the same 1% concentrated into a two-second burst is a hole in a sentence. Video behaves the same but with pictures. Compressed video frames depend on their neighbors. So one lost packet can smear artifacts across the screen until the next full frame resets the image.


The rule of thumb that falls out: TCP traffic converts loss into slowness; real-time traffic converts loss into damage. Slowness scales gently — damage does not, which is why the voice thresholds below are stricter than everything else.


Working thresholds by path type


With the mechanisms understood, here are working numbers you can defend in a meeting. Treat them as engineering rules of thumb — starting points to tune against your own baseline, not laws of physics:


Path typeExcellentAcceptableInvestigate
Wired LAN0% measuredRare momentary blips, well under 0.1%Any sustained measurable loss
Wi-Fi (post-retry)Under 0.5%0.5–2% at busy timesOver 2% sustained
WAN / internet pathUnder 0.1%Brief excursions up to about 1%Over 1–2% sustained
Voice / video path (end to end)Under 0.5%0.5–1%, scattered not burstyOver 1%, or any bursts; over 3% means bad calls

Wired LAN: the standard really is (effectively) zero

A modern switched LAN is full duplex end to end — no collisions, no shared medium, links that are either healthy or faulty. There is no natural background loss to tolerate, so measured loss inside your own LAN always has a findable cause. It could be a damaged cable, a duplex mismatch, or a congested uplink. That is why the "investigate" bar sits at any sustained measurable loss, not at some percentage. On the LAN, loss is not a level to manage but a defect to locate. The causes guide is the map for that hunt.


Wi-Fi: judged on a curve, for physics reasons

Wireless gets a gentler scale because the radio layer retransmits lost frames automatically before your tools ever see them. The loss you measure is what survived several retries, and the retries themselves already cost latency. A little post-retry loss at busy times is the shared airtime showing through. Sustained loss above a couple of percent means the radio environment is genuinely struggling and worth engineering work. The full story of why wireless "loss" is a different animal lives in the Wi-Fi loss guide.


WAN and internet paths: small is normal, sustained is not

Long shared paths cross dozens of routers you do not control, and momentary buffer overflow somewhere along the way is routine. That explains the tolerance for brief excursions. But remember the square-root math. Sustained loss above about 1% on a long-RTT path does not make TCP a little slower. It caps throughput at a small fraction of the line rate. And your users will feel it as "the internet is slow" every afternoon. Sustained loss on a WAN path is worth investigating even when each individual reading looks small.


Voice: strict, and bursty loss counts double

The voice numbers are the strictest because nothing hides the damage. Under half a percent, concealment keeps calls clean. Around 1%, quality depends heavily on distribution — scattered single losses conceal well; bursts do not. By 3%, enough syllables are being patched or dropped that ordinary users describe the call as bad. When you monitor a voice path, if your tools can report burstiness or consecutive-loss events alongside the percentage, treat that as first-class information rather than trivia.


The measurement window problem


Here is the part almost everyone skips. Skipping it invalidates more loss conversations than any other mistake: a loss percentage means nothing without the time window it was measured over. Consider a network that runs perfectly clean all day except for one horrible 30-second episode of 5% loss during the morning all-hands call. Averaged over the day, that episode works out to roughly 0.002% loss — a number that would win awards. Averaged over the 30 seconds that mattered, it was 5% — a number that wrecks a meeting. Same network, same day, same data: one window says "flawless," the other says "broken."


Neither number is a lie; they answer different questions. Long windows (hours, days) answer "is this path chronically degraded?" and are what you trend and report. Short windows (seconds to a few minutes) answer "is it bad right now?" and are what users actually experience — nobody experiences a daily average. A monitoring setup needs both. A written threshold needs its window attached or it is not a threshold at all. "loss under 1%" is incomplete. "loss under 1% over any 10-minute window" is something you can measure, alert on, and hold a provider to.


Window honesty also applies when comparing numbers from different sources. A monitoring dashboard averaging five-minute polls, a ping test that ran for ninety seconds, and a provider's monthly report are all speaking different dialects of "percent." Stapling them together produces arguments, not diagnoses. When two loss figures disagree, check the windows before you check the network — more than once, the "discrepancy" turns out to be arithmetic.


Tip: make "over what window?" a reflex. When a colleague reports "2% packet loss," the first question is how long the measurement ran and when. A 20-ping spot check and an all-day monitor produce utterly different meanings for the same percentage. How to measure so the number deserves trust is its own craft — covered in the measurement guide.


When to alert: sustained versus momentary


Thresholds become useful the day they drive alerts — and dangerous the day they drive too many. The core distinction to build in is momentary versus sustained. A single bad sample — one minute showing 3% loss, surrounded by clean minutes — is a microburst, a backup kicking off, or measurement noise. If you page yourself for every one, you will be ignoring your own alerts within a month. Loss that persists across consecutive windows is a condition, and conditions are what deserve humans.


The pattern that works in practice is two rules per monitored path, aimed at two different failure shapes:


  • A fast rule for outage-grade loss: high threshold, short persistence — for example, alert when loss exceeds 10% for three consecutive minutes. This catches the failing link and the dying interface within minutes, without firing on a single rough sample.
  • A slow rule for chronic degradation: use a low threshold and long persistence. For example, warn when loss stays above 1% for 15 minutes, or when the daily average creeps above your baseline. This catches the slowly corroding cable and the afternoon congestion pattern that no single dramatic spike ever announces.

Route the two differently: the fast rule can interrupt your day; the slow rule should open a ticket for working hours. And tie severity to traffic type. A voice path warning at 1% is reasonable, while the same 1% on a bulk-data path might merit only a trend line. An alert you routinely ignore is worse than none, because it teaches you to ignore the next one.


A worked example: thresholds for a small office


Time to assemble the pieces. Take a concrete office: forty people, VoIP desk phones, a local file server, Wi-Fi everywhere, one internet line, and cloud apps that matter. Here is a defensible threshold sheet, built step by step:


  • Step 1 — pick the paths to watch. Four cover the experience. There is the internet path (edge to a reliable outside target) and the LAN core (desktop switch uplinks to the server). There is the Wi-Fi path (a wireless client's route to the LAN). And there is the voice path (phones to wherever the phone system lives — the internet path again, if it is hosted).
  • Step 2 — pick the windows. Sample every minute; evaluate rules over rolling windows. One-minute samples are fine-grained enough to catch real episodes without drowning you in noise.
  • Step 3 — set the internet rules. Fast: over 10% loss for 3 consecutive minutes → alert (the line is failing). Slow: over 1% sustained for 15 minutes → ticket (congestion or a degrading circuit; if it recurs daily, the time-of-day pattern is your evidence).
  • Step 4 — set the LAN rule. Any measurable loss sustained 10 minutes on a LAN path → ticket, because the LAN standard is zero. Pair it with interface-counter checks on the uplinks — rising discards tell you it is congestion rather than damage.
  • Step 5 — set the voice and Wi-Fi rules. Voice: warn at 1% over 5 minutes, alert at 3% — calls are already suffering at three. Wi-Fi: warn at 2% sustained 15 minutes, and resist tightening it until you have a baseline; wireless needs one before its numbers mean anything.
  • Step 6 — baseline, then tune. Run two weeks and expect to adjust. Perhaps the Wi-Fi warning fires every lunchtime (loosen it, or fix the airtime problem it found). And perhaps the internet slow rule catches a real nightly pattern (evidence for the ISP conversation). Write the final numbers down with their windows, because six months from now "1%" without context will be useless to everyone, including you.

Where to go next


You now have what the one-line answers never provide: the reasoning. TCP turns loss into slowness that scales with round-trip time. Real-time traffic turns loss into damage that concealment only partly hides. Every threshold needs a window attached. And alerting should distinguish the momentary from the sustained. The natural next step is learning to measure loss accurately so your numbers deserve the thresholds. Then trace any loss you confirm back to its source with the causes tour. The full set of guides lives in the Packet Loss topic.


Put this into practice with SysPeep Monitor

Thresholds only help if something is watching them around the clock. SysPeep Monitor polls your links and interface counters. It evaluates sustained-versus-momentary rules the way this guide describes. Its alert rules and dependencies are designed to keep the noise down. So a warning actually means something. The free trial installs on any Windows machine in minutes.

Download Free TrialLearn About SysPeep