MTR: Continuous Path Testing Done Right
Traceroute has a dirty secret: three probes per hop is an anecdote. Run it during a problem and you get one blurry snapshot — maybe the bad moment, maybe not. MTR is what happens when traceroute refuses to stop. It discovers the path the same way, then keeps probing every hop, over and over. It builds live statistics — loss percentage, average, best, worst, jitter — for every router between you and the destination. Hundreds of samples per hop instead of three. Anecdotes become evidence.
That power comes with the same trap that catches traceroute readers, magnified: MTR shows loss numbers at every hop. And most of those numbers do not mean what they appear to mean. This guide covers installing the tool and running it in the modes that actually help. And — the heart of it — it covers reading the loss column correctly. That is the single most valuable skill this whole family of tools has to teach.
In this guide
- What MTR is: traceroute that keeps going
- Getting it on your machine (and the Windows situation)
- Running it usefully: the flags that matter
- The core skill: reading loss down the hop table
- Latency patterns worth recognizing
- Building evidence a provider will accept
- The overnight stakeout — and its limits
- Reading the report: a pocket table
What MTR is: traceroute that keeps going
MTR (the name comes from "My Traceroute") uses the same trick as traceroute: probes with deliberately short TTL hop counters. Those provoke each router along the path into revealing itself with a "time exceeded" complaint. If that trick is new to you, read reading traceroute first — everything there about quiet routers, lazy control planes, and invisible return paths applies here word for word.
The difference is persistence. Where traceroute asks each hop three times and quits, MTR cycles endlessly. It sends one probe to every hop, then again, then again, once per second by default. After five minutes you have three hundred samples per hop, and questions that traceroute cannot touch become answerable: Is that loss constant or bursty? Is the latency jump always there or only at 3 p.m.? Did the path change? How often, how much, since when — the questions tickets are made of.
Sample size is not a luxury here; it is the entire point. Three probes cannot distinguish 1% loss from 0% — a hundred can barely resolve 1% steps, and fractions of a percent take thousands of samples. MTR is the tool in this family that can actually accumulate those samples per hop while you watch. (The statistics of loss measurement get a fuller treatment in measuring packet loss.)
Getting it on your machine (and the Windows situation)
On Linux, MTR is one package away. Use sudo apt install mtr-tiny on Debian and Ubuntu (or mtr for the graphical build). Use sudo dnf install mtr on RHEL-family systems. On macOS, brew install mtr does it. Because MTR crafts raw probe packets, it typically needs root: run it with sudo.
Windows deserves an honest paragraph. There is no MTR in the box. The built-in cousin is pathping, which shares the spirit in batch form. It traces the route, then quietly sends around a hundred probes to every hop — taking several minutes. It prints one final per-hop loss table. No live view, no flag playground, ICMP only, but it is on every Windows machine you will ever touch. And its loss table is read with exactly the rules this guide teaches:
C:\> pathping -n 203.0.113.80 # in a hurry: fewer probes per hop C:\> pathping -n -q 25 203.0.113.80
Windows ports of MTR also exist — WinMTR is the commonly encountered one — giving you the live cycling view. The usual software hygiene applies: fetch tools like these from their official project source, and clear them with whoever approves software in your shop. For everything below, the Linux/macOS command line is shown; the reading skills transfer to pathping and the ports unchanged.
Running it usefully: the flags that matter
MTR has two personalities. Interactive mode — just sudo mtr 203.0.113.80 — fills the terminal with a live table that updates every second. It is hypnotic, and genuinely useful when you want to watch a problem happen. Wiggle the suspect cable, start the big backup job, or join the test call. See which hop's numbers move in real time. But the mode that earns its keep is report mode: run a fixed number of cycles, print one clean table, exit. Fixed runs are comparable and shareable — the currency of tickets:
# 100 cycles, wide output, no name lookups, then print and exit sudo mtr -rwn -c 100 203.0.113.80 # same, but show hostnames AND addresses sudo mtr -rwb -c 100 203.0.113.80 # gentler pace: one cycle every 2 seconds sudo mtr -rwn -c 100 -i 2 203.0.113.80
Here are the flags, decoded. The flag -r is report mode. The flag -w is wide, so hostnames are not truncated. The flag -n skips DNS lookups (faster, less clutter — use -b instead when you want names and numbers together). The flag -c 100 sets the cycle count. A hundred is a solid default, enough samples to make the loss column meaningful. The flag -i 2 stretches the interval between cycles. One cycle per second is polite. Going much faster mostly triggers the rate limiting you are about to read about, and sub-second intervals need root anyway. MTR can also switch probe flavors like its parent — --tcp or --udp when firewalls treat plain ICMP coldly.
A report looks like this — and the columns map straight onto ping's summary numbers, per hop:
HOST: adminpc Loss% Snt Last Avg Best Wrst StDev 1.|-- 192.168.1.1 0.0% 100 0.4 0.5 0.3 1.9 0.2 2.|-- 100.64.0.1 0.0% 100 11.8 12.1 10.9 19.4 1.1 3.|-- 198.51.100.17 0.0% 100 12.9 13.0 12.1 20.2 1.2 4.|-- 198.51.100.62 0.0% 100 13.4 13.6 12.8 15.1 0.4 5.|-- 203.0.113.9 0.0% 100 14.2 14.5 13.4 22.7 1.3 6.|-- 203.0.113.80 0.0% 100 14.8 15.0 13.9 21.5 1.2
Snt is probes sent to that hop; Loss% is the fraction unanswered. Best is the floor — the distance signal. Avg is the headline, Wrst the worst single moment, and StDev the spread — the jitter needle. A calm path shows Best and Avg close together and StDev tiny, exactly like this one.
The core skill: reading loss down the hop table
Here is the lesson that repays the whole guide. MTR's loss column tempts you to find the first non-zero number and declare victory. Resist. Routers answer probes with their busy little management CPUs as a courtesy, and many deliberately rate-limit those answers — while forwarding actual traffic flawlessly in hardware. A hop can therefore "lose" most of your probes and be completely healthy. The question is never "where is loss shown?" but "does the loss continue to the destination?" Compare these two reports of the same path.
Report one — the false alarm:
HOST: adminpc Loss% Snt Last Avg Best Wrst StDev 1.|-- 192.168.1.1 0.0% 100 0.4 0.5 0.3 1.9 0.2 2.|-- 100.64.0.1 0.0% 100 11.8 12.1 10.9 19.4 1.1 3.|-- 198.51.100.17 0.0% 100 12.9 13.0 12.1 20.2 1.2 4.|-- 198.51.100.62 60.0% 100 13.4 13.6 12.8 15.1 0.4 5.|-- 203.0.113.9 0.0% 100 14.2 14.5 13.4 22.7 1.3 6.|-- 203.0.113.80 0.0% 100 14.8 15.0 13.9 21.5 1.2
Sixty percent loss at hop 4 looks like a disaster — until you read on. Hops 5 and 6 show 0.0%, and every probe that reached them traveled through hop 4. If hop 4 were genuinely dropping six packets in ten, the hops behind it could not possibly be clean. Verdict: hop 4's router is rate-limiting its own replies and forwarding everything. Users feel nothing. No ticket. This single pattern — scary number at one hop, silence downstream — is the most misdiagnosed sight in all of network troubleshooting.
Report two — the real thing:
HOST: adminpc Loss% Snt Last Avg Best Wrst StDev 1.|-- 192.168.1.1 0.0% 100 0.4 0.5 0.3 1.9 0.2 2.|-- 100.64.0.1 0.0% 100 11.8 12.1 10.9 19.4 1.1 3.|-- 198.51.100.17 0.0% 100 12.9 13.0 12.1 20.2 1.2 4.|-- 198.51.100.62 4.0% 100 13.4 38.6 12.8 95.1 21.3 5.|-- 203.0.113.9 5.0% 100 14.2 39.5 13.4 99.7 22.0 6.|-- 203.0.113.80 4.0% 100 14.8 40.0 13.9 98.5 21.8
A smaller number — 4% — but it appears at hop 4 and continues, at similar magnitude, all the way to the destination. That is real loss entering the path around hop 4. The swollen Avg, Wrst, and StDev in the same rows tell the same story: a congested link both dropping and delaying. Users feel this one — stalls, retransmissions, robot voices on calls. This report is a ticket, and the sentence that goes with it is: "loss begins at hop 4 and persists to the destination."
Two footnotes keep you honest. First, the last row is the ground truth: it is the only row that describes what end-to-end traffic experiences. There is a mechanical reason to trust it more, too. Middle hops answer with grudging, rate-limited "time exceeded" complaints, while the destination sends ordinary echo replies. So the destination's numbers are far less polluted by router etiquette. Middle rows are diagnostic color that tell you where trouble starts, never whether trouble exists. If the last row is clean over a decent sample, users are fine no matter how alarming the middle looks. (If the destination filters probes entirely, the last answering network stands in as your best proxy — noted as such, with appropriate humility.) Second, compare like with like: small differences between hops (4% versus 5% on 100 probes) are sampling noise, not a finding. And if loss starts at hop 1 or 2 and persists, stop drafting the ISP ticket — the problem is in your own building, most often the Wi-Fi.
Latency patterns worth recognizing
With hundreds of samples per hop, the latency columns develop readable shapes:
- The gradual stairstep. Best climbs in small, steady increments down the table, with one or two big risers where long-haul geography lives. This is a healthy path wearing its distances honestly.
- One big step that persists. Avg and Best jump at a hop and stay elevated for every hop below it. Real added delay from that link onward: an ocean if it is constant, a congested link if it swells at busy hours. The before/after comparison in the evidence section settles which.
- Wrst far above Avg, rhythmically. Avg says 15 ms, Wrst says 180 ms, and repeated runs show it again and again: periodic queue spikes — congestion with a pulse. Capture reports at different times of day and the rhythm usually maps onto somebody's busy hours. That is exactly the correlation a provider (or your own bandwidth graph) can act on.
- StDev growing down the table. Jitter accumulating along the path. Interactive users and VoIP feel this before anyone else; the measurement side of that story is covered in measuring latency.
- The path itself changing mid-run. Hops renumber, new addresses appear, question marks flicker. Routing moved underneath you — sometimes a provider rerouting around trouble, sometimes load balancing showing its parallel roads. A long run that straddles a path change explains many a "it got slow at exactly 2:15" mystery.
All the traceroute caveats still apply: a spike or loss figure confined to one middle hop with a clean tail is control-plane laziness. And every number includes an invisible return leg. MTR does not repeal those rules; it just gives you enough samples to apply them with confidence.
Building evidence a provider will accept
MTR reports are the standard attachment for path complaints — when they are captured so a stranger can trust them. The checklist:
- Fixed counts, identical flags. Every capture the same way —
-rwn -c 100— so tables are directly comparable. - Good times and bad times. One report during the problem and one during a quiet hour. The contrast is the evidence; a single bad report proves almost nothing.
- Timestamps and time zone. Note when each report ran (prefix the run with
dateand capture both). "Around lunchtime" is not data. - Two vantage points if you can get them. The same destination from the office and from a phone hotspot or home line. One dirty path plus one clean one pins the problem to a network rather than to the destination.
- Name the pattern in one sentence. "Loss begins at hop 4 and persists to the destination; clean at 7 a.m., 4–5% at 2 p.m." A sentence like that — which quietly demonstrates you know rate-limited middle hops are not loss — moves a ticket from the form-letter pile to an engineer's queue.
The overnight stakeout — and its limits
Intermittent problems rarely perform on demand. For those, run MTR as a stakeout: scheduled report snapshots appended to a file, so tomorrow you can scroll to the moment things went wrong. A small loop does it — a 60-cycle report roughly every five minutes:
while true; do date >> path-watch.log mtr -rwn -c 60 203.0.113.80 >> path-watch.log sleep 240 done
Leave it running in a spare terminal (or under screen or tmux so it survives you logging out). On Windows, the same idea is a Task Scheduler job appending pathping output to a file on a schedule. When the complaint lands — "the video call fell apart around 2:40" — you search the log for the nearest timestamps. Then you read the before, during, and after tables. It is a genuinely effective poor-man's monitor, and a polite one: sixty probes every five minutes is a whisper of traffic. And the text file grows by mere kilobytes an hour.
And it is worth being honest about what it is not. A log file does not alert you, graph trends, or watch more than one target. It does not survive the laptop going to sleep, or notice at 3 a.m. so you can sleep. Hand tools investigate; monitoring systems watch — continuously, against baselines, with alerts when patterns change. The graphing half of that idea, latency history you can see at a glance, is the subject of Smokeping-style latency graphing. When a stakeout starts feeling like a permanent fixture, that is the signal you have outgrown the loop.
Tip: capture one -c 100 report to each destination you truly depend on while everything works, and file it. The day something breaks, that baseline turns "it feels slow" into "hop 5's Avg tripled and StDev went from 1 to 20" — a sentence that gets action.
Reading the report: a pocket table
| Pattern in the table | What it means | What to do |
|---|---|---|
| Big loss at one middle hop, 0% below it | Router rate-limiting its own replies; forwarding fine | Nothing. Not a problem |
| Loss starts at hop N, persists to the last row | Real loss entering the path near hop N | Capture good/bad reports; escalate with the persist sentence |
| Loss on every row including hops 1–2 | Trouble in your own building or access link | Check LAN, Wi-Fi, and the local router first |
| Last row clean over hundreds of probes | Path is fine end to end | Look at the service, not the network |
| One persistent latency step down the table | Distance, or a congested long-haul link | Compare quiet-hour vs busy-hour reports |
| Wrst and StDev ballooning at rhythm | Periodic congestion — queues with a pulse | Correlate with time of day and utilization |
| Hops flapping between two routers | Load-balanced parallel paths | Normal; note it, compare their numbers |
Where to go next
MTR is the bridge tool of this family: traceroute's map plus ping's statistics, in one screen. Master the persist rule for loss, trust the last row, and capture fixed-count reports at contrasting times. Then you are already reading paths better than most of the internet. For the foundations underneath, see what ping actually tells you and reading traceroute. For the always-on version of what your stakeout loop was reaching toward, continue to Smokeping-style latency graphing and the rest of Ping, Traceroute & Friends.
Put this into practice with SysPeep Monitor
An MTR stakeout is you doing a monitoring system's job by hand. SysPeep Monitor does the watching part properly. It auto-discovers your devices and collects health and performance data over SNMP, WMI, and SSH. It graphs the data on live dashboards and alerts on thresholds with dependency rules that keep the noise down. So when you reach for mtr, you already know when it started and what else broke. The free trial installs on one Windows machine.
Download Free TrialLearn About SysPeep