Measuring Loss Accurately: Why Ping Undercounts and What to Use
Here is a scene that plays out in IT rooms every day. A user reports terrible call quality. The admin runs a quick ping — twenty packets, 0% loss — declares the network healthy, and closes the ticket. The user keeps suffering, because the network genuinely is dropping packets; the measurement just could not see it. Nobody lied. The tool was simply asked a question it cannot answer.
Packet loss investigations go wrong at the measurement step more often than at any other. That is because the default tool — ping — has blind spots that are invisible until someone points them out. It samples far too little to see small loss rates. It gets deprioritized by the very routers it is testing. And it probes one path with one packet size on one metronomic rhythm while your real traffic does none of those things.
This guide fixes your measurements. You will learn exactly why ping undercounts (and sometimes overcounts). You will learn the simple sample-size math that tells you how many probes a trustworthy number needs. And you will learn the tools that answer the questions ping cannot: interface counters, mtr, and iperf3 — each with real commands and annotated output. At the end, a decision guide matches each method to the question it is best at.
In this guide
- What ping actually measures
- Ping's three blind spots
- Fix 1: bigger, longer samples (the math)
- Fix 2: interface counters — the ground truth
- Fix 3: mtr — reading loss per hop
- Fix 4: iperf3 — controlled loss testing
- Spot checks versus continuous monitoring
- The decision guide
What ping actually measures
Ping sends a small test message (an ICMP echo request) to a target, which is supposed to answer with an echo reply. For each probe, either an answer comes back or it does not. The loss percentage at the end is just "unanswered probes divided by probes sent." Simple — but notice what a single "lost" ping actually means. The request or the reply died somewhere along the outbound path, inside the target, or along the return path. Ping cannot tell you which, and it measures both directions mashed together. Keep that in mind for everything that follows: a ping statistic is a round-trip verdict on a one-way question.
None of this makes ping useless — it is fast, universal, and a fine first probe. The trouble starts when a quick ping is treated as a loss meter. It fails at that job in three distinct ways, and each failure points to a better tool.
Ping's three blind spots
Blind spot 1: the sample is far too small
Loss rates that wreck real applications are small numbers. A voice path is already degrading at 1% loss — one packet in a hundred. And a long TCP transfer suffers at rates well below that. Now do the math on a casual test: with 20 pings, a single lost packet reads as 5%. The counter cannot even display a value between 0% and 5%. With 100 pings you can at best see whole percents. And catching a rare event takes luck. Suppose the true loss rate is 0.1% (one packet in a thousand). Then a 100-ping test will, about nine times out of ten, lose nothing at all and proudly report 0% loss. The network was dropping packets the entire time; the sample was simply too small to contain one.
Blind spot 2: routers answer ping with their spare time
Generating an ICMP reply is extra work a router's processor does as a favor, and routers are built to forward first and chat later. Most deprioritize or rate-limit their ping responses on purpose. So a perfectly healthy router that is busy forwarding at full speed may ignore a chunk of your probes. That shows you scary "loss" that its transit traffic never experiences. This matters most when you ping intermediate hops (it is the great trap of traceroute-style tools, dissected in the mtr section below). Some firewalls go further and drop ICMP entirely, producing a permanent 100% "loss" to a service that is working fine. The general lesson: loss reported by a middle device is a claim about its manners, not about your path.
Blind spot 3: one path, one size, one rhythm
Your ping probes are identical little creatures: the same small size, the same source and destination, launched once per second like a metronome. Real traffic is none of that, and the differences hide loss. Networks often split traffic across parallel links by hashing each flow's addresses. So all your pings ride one member link while the faulty parallel link eats other people's packets. Marginal cabling corrupts big frames more readily than small ones — a 64-byte ping sails through where a full-size 1,500-byte frame gets mangled. And a once-per-second rhythm can miss loss that comes in millisecond microbursts between the ticks. A clean ping means "this size, this path, these instants were fine" — a much weaker statement than it appears.
Fix 1: bigger, longer samples — with the math made simple
The sample-size problem has a one-line cure: send enough probes that the loss you care about would show up many times, not zero-or-once. The rule of thumb: expect at least ten lost packets in your sample before you trust the percentage. Since expected losses are simply rate × count, the required count is ten divided by the rate you want to detect:
| Loss rate you need to see | Probes required (rule of thumb) | At 5 probes/second |
|---|---|---|
| 1% (voice-wrecking) | 1,000 | About 3.5 minutes |
| 0.5% | 2,000 | About 7 minutes |
| 0.1% (TCP-throttling on long paths) | 10,000 | About 33 minutes |
| 0.01% | 100,000 | Over 5 hours |
This is why a serious ping test looks like this on Linux or macOS — a thousand probes, five per second, near full size:
ping -c 1000 -i 0.2 -s 1400 192.0.2.50
# -c 1000 send 1,000 probes, not 4 or 20
# -i 0.2 one probe every 0.2 s (the fastest most systems
# allow without administrator rights)
# -s 1400 1,400-byte payload -- big frames find marginal
# cables that tiny ones slip past
--- 192.0.2.50 ping statistics ---
1000 packets transmitted, 993 received, 0.7% packet loss
^^^^
7 lost of 1,000: believable, but near
the edge -- run it again to confirm
On Windows the equivalent is ping -n 1000 -l 1400 192.0.2.50 (it always probes about once per second, so budget more time). Two readings of the summary line matter: the percentage, and the raw counts behind it. A result of 7 lost of 1,000 is a measurement. A result of 1 lost of 20 is a coin flip. And note the table's last row: verifying that a path is clean to the 0.01% level takes hours of probing. That is your first hint that one-off tests eventually lose to continuous measurement.
One refinement beats even bigger samples: spread them. Loss is rarely uniform across the day. So three 1,000-probe runs — morning, midday, and evening — tell you more than one 3,000-probe marathon. They can catch a pattern as well as a rate. If the morning run is clean and the evening run is ugly, you have not merely measured the loss. In that case, you have dated it. And dating a problem is halfway to explaining it.
Fix 2: interface counters — where ping guesses, counters count
Everything ping-shaped shares one weakness: it samples. Send test packets, hope they experience what real traffic experiences, extrapolate. But for any specific link you control, there is a source of truth with no sampling error at all. That source is the interface counters on the devices at each end. Every port keeps lifetime totals of frames that arrived damaged (errors) and frames it had to throw away (discards). Those totals are counted across every frame that crossed the port, including all the real traffic, at every instant. Ping estimates; counters are the actual body count.
On a Linux host, one command shows them:
ip -s link show eth0
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 ...
RX: bytes packets errors dropped missed mcast
8123456789 9876543 213 448 0 12345
^^^ ^^^
damaged on arrival thrown away undamaged
(cable/NIC territory) (buffer/overload territory)
TX: bytes packets errors dropped carrier collsns
5432109876 7654321 0 0 0 0
On a switch, the same information appears in show interface-style output, and — crucially for monitoring — it is all readable remotely over SNMP. The counters' limits are the mirror image of ping's: they are per-link, not end-to-end. So they tell you everything about one segment and nothing about the rest of the path. And they are lifetime totals, so a single glance means little — what matters is growth between two readings. Interpreting each counter family, and the combinations that point to specific faults, is a craft of its own. And the interface counters guide is the deep dive. For this guide's purpose, remember the division of labor — ping tells you a path has a problem; counters tell you which link is committing it.
Fix 3: mtr — reading loss patterns hop by hop
mtr (My Traceroute) combines traceroute and ping. It discovers each router along the path, then probes all of them continuously, tabulating loss and latency per hop. It is the tool for the question "where along the path does loss begin?" There is one enormous caveat you already know from blind spot 2. Intermediate hops answer with their spare time, so mid-path numbers can lie. The skill of mtr is reading the pattern, not any single row. A serious run uses report mode and plenty of cycles:
mtr -rwc 300 203.0.113.80 # -r report mode (run, then print a summary table) # -w wide output with full hostnames # -c 300 three hundred probe rounds -- sample size again HOST: adminpc Loss% Snt Last Avg Best Wrst 1.|-- 192.168.1.1 0.0% 300 0.4 0.5 0.3 2.1 2.|-- 100.64.17.1 0.0% 300 8.2 9.1 7.4 31.0 3.|-- 198.51.100.9 28.3% 300 9.0 9.4 8.1 22.7 4.|-- 198.51.100.66 0.0% 300 12.1 12.6 11.0 40.2 5.|-- 203.0.113.5 2.3% 300 19.8 20.4 18.9 55.1 6.|-- 203.0.113.80 2.7% 300 20.1 20.9 19.2 58.4
Two patterns in that output, and they are the two patterns to learn. Hop 3 shows 28% loss — but every hop after it is clean or nearly so. That is impossible for real loss. Packets that die at hop 3 can never reach hops 4 through 6. So if hop 3 were truly eating a quarter of the traffic, every later row would show at least that much loss too. Verdict: hop 3 is just rate-limiting its ping replies while forwarding perfectly. Ignore it.
Hops 5 and 6 show the real thing: loss that starts mid-path and persists to the final hop. The destination row is the only row measuring true end-to-end loss. When it shows roughly the same figure as some earlier hop onward, that earlier hop is your first suspect for where the trouble begins. Here, that means somewhere between hop 4 and hop 5. The reading rules compress to: the last hop is the truth; a middle hop matters only if its loss continues all the way down. Note one honest limitation in your findings. Mtr sees the forward path. Return-path problems show up in its numbers without being visible in its hop list. When stakes are high, run it from both ends.
Fix 4: iperf3 — controlled loss testing between two points
Sometimes the question is not "is there loss right now?" but "will this path carry the load we plan to put on it?" Can the site-to-site VPN handle thirty voice calls? Will the new backup window saturate the uplink? For that you need a load you control, and that is iperf3. Run it as a server on one end, and aim a client at it from the other. It generates precise traffic and counts exactly what arrives. The crucial flag is -u for UDP mode. In its default TCP mode, retransmission hides loss (you would measure reduced throughput instead). In UDP mode every lost datagram is counted by name:
# far end (e.g. the branch office server):
iperf3 -s
# near end -- send 20 Mbit/s of UDP for 30 seconds:
iperf3 -c 192.0.2.10 -u -b 20M -t 30
[ ID] Interval Transfer Bitrate Jitter Lost/Total
[ 5] 0.00-30.00 sec 71.5 MB 20.0 Mbits/s 0.21 ms 183/51840 (0.35%)
^^^^ ^^^^^^^^^
bonus metric: the answer: 183 of
arrival-time 51,840 datagrams
wobble (jitter) died en route
Read the far (server) end's report for the authoritative received count. Choose -b to match the load you actually plan to run. Testing a link at full line rate proves only that a flooded link drops packets, which you knew. The jitter column is a welcome bonus: real-time traffic cares about it nearly as much as loss (the Network Latency guides pick that thread up).
Warning: iperf3 is not a passive observer. It generates the traffic it measures, at whatever rate you request. UDP mode keeps sending regardless of what it squeezes out of the way. Run meaningful rates outside business hours or in maintenance windows. Start modest. And never aim a high-bandwidth test across a production WAN link at 10 AM to "see what happens." You will see what happens.
Spot checks versus continuous monitoring
Every method so far shares a quiet assumption: that the loss will politely occur while you are testing. Real loss rarely cooperates. It happens during the 9 AM logon storm, during the nightly backup, or when the microwave runs. It happens when the ISP's neighborhood segment saturates at 8 PM. And the admin who tests at 3 PM finds a spotless network every single time. Intermittent loss is its own discipline with its own workflow (covered in this topic). But the measurement lesson belongs here: a spot check tells you about a moment; only continuous measurement tells you about a network.
Continuous measurement also gives you the thing every threshold quietly depends on: a baseline. A reading of 2% loss is an emergency on a path that has run clean for a year. It is business as usual on a path that has wobbled around 2% forever. Only stored history can tell you which network you are standing in.
The machinery is less grand than it sounds. A monitoring system polls every interface's error and discard counters every minute or two. Alongside that runs a steady low-rate ping series to a few key targets. All those measurements are stored with timestamps. The stored history is what transforms troubleshooting. "Loss on the WAN spikes every weekday 8–9 PM" is a diagnosis-shaped fact no spot check can produce. The counter history pinpoints which link misbehaved at 2:14 AM without anyone being awake for it. The one-off tools in this guide remain essential for investigation; the monitoring provides the detection and the timeline. (For assembling a broader measurement toolkit on a budget, see measuring the core metrics with free tools.)
The decision guide: which method for which question
| Your actual question | Best method | Why |
|---|---|---|
| Is this path losing packets right now? | Long ping: 1,000+ probes, large size | Enough sample to see small rates; large frames catch marginal links |
| Which link on the path is at fault? | Interface counters at each hop you control | Counts every real frame — no sampling, per-link verdicts |
| Is the loss inside my network or beyond the edge? | mtr from a machine at the edge | Shows where loss begins — if it persists to the last hop |
| Can this path carry the planned real-time load? | iperf3 UDP at the planned bitrate | Controlled load, exact lost-datagram count, jitter included |
| When does it happen, and is it getting worse? | Continuous monitoring of counters + ping series | Only stored history reveals patterns and trends |
| Is the ISP's segment the problem? | mtr runs + timestamped long pings, collected over days | Evidence a provider will accept beats a single angry screenshot |
Where to go next
The habit to take away: match the tool to the question, and distrust any loss figure that arrives without a sample size and a time window attached. A twenty-ping test can rule loss in (if it shows loss, believe it) but never rule it out. From here, one natural next guide is reading interface counters correctly — the ground-truth skill this guide only opened. Another is how much loss is acceptable, which turns your new, trustworthy numbers into thresholds worth acting on. The rest of the toolkit lives in the Packet Loss topic.
Put this into practice with SysPeep Monitor
The hands-on tools in this guide answer today's question; monitoring answers next month's. SysPeep Monitor discovers your switches and routers. It polls their interface error and discard counters over SNMP around the clock and alerts you when loss fingerprints start growing. So the 2 AM problem is on a graph by morning. The free trial installs on any Windows machine in minutes.
Download Free TrialLearn About SysPeep