
Ping, traceroute and mtr: what each one can and cannot tell you
Ping vs traceroute is really a question about which problem you have. Ping asks one host whether it is there and how long a round trip takes. Traceroute lists the routers on the path to that host, sampled once. mtr does both at the same time and keeps going, so you get loss and latency for every hop over minutes rather than a single snapshot. Use ping for "is it reachable and how fast", traceroute for "which way does the traffic go", and mtr for "where along the path does it go wrong". For anything that needs a ticket to a provider, send an mtr report.
Ping vs traceroute vs mtr at a glance
| ping | traceroute | mtr | |
|---|---|---|---|
| Sends | ICMP echo requests to one address | Probes with rising TTL (UDP, ICMP or TCP) | Probes with rising TTL, repeated every second |
| Measures | Round-trip time and loss to the target | The routers on the path and three round trips to each | Loss, latency and jitter to every hop over time |
| Runs | Until you stop it, or -c packets |
Once | Until you stop it, or -c cycles in report mode |
| Good for | Reachability, latency baselines, MTU tests | Seeing the path and spotting a routing change | Locating loss and latency on the path |
| Cannot tell you | Where a problem is, or whether the service works | Whether a problem is steady or brief | The return path, or what users' traffic gets |
All three share one limit: they measure how routers and hosts treat probe packets, which is not always how they treat your users' traffic. Keep that in mind and most misreadings go away.
ping: one host, round trip and loss
# Linux and macOS: 20 echoes, then a summary
ping -c 20 203.0.113.10
# Windows
ping -n 20 203.0.113.10
The summary is the part to read. On Linux (iputils) it looks like this:
--- 203.0.113.10 ping statistics ---
20 packets transmitted, 20 received, 0% packet loss, time 19027ms
rtt min/avg/max/mdev = 11.204/11.871/14.310/0.602 ms
Minimum is your floor, close to the true path latency. Average and max show queuing on the way. mdev is the
spread; on a quiet wired path it stays well under a millisecond, and a value several times higher than your
usual one means congestion or Wi-Fi somewhere. macOS prints the same numbers as round-trip min/avg/max/stddev.
Sample size matters more than people expect. With 20 echoes, a single lost packet reads as 5 percent loss. If you want to tell 0.5 percent loss from 1 percent, you need hundreds of samples, which is why a quick ping that "lost one" proves very little.
Two Linux flags help when you leave ping running during an incident: -D prints a Unix timestamp on every line
so you can line replies up with logs, and -O prints a line for each reply that never came, which makes gaps
visible instead of silent.
ping -D -O 203.0.113.10 | tee /tmp/ping-203.0.113.10.log
Ping is also the quickest MTU test. 1472 bytes of data plus 8 bytes of ICMP header plus 20 bytes of IP header is exactly 1500, so this checks whether a full-size packet gets through with fragmentation forbidden:
ping -M do -s 1472 -c 3 203.0.113.10 # Linux
ping -D -s 1472 -c 3 203.0.113.10 # macOS
ping -f -l 1472 -n 3 203.0.113.10 # Windows
If 1472 fails and 1372 works, something on the path, usually a tunnel or PPPoE link, has a smaller MTU. That one test explains a whole class of "small pages load, big downloads hang" tickets.
What ping cannot do is say where a problem is, and a reply proves only that a kernel answered, not that a service works. The second point is the subject of ping is not monitoring, so I will not repeat it here.
traceroute: the path, sampled once
Traceroute sends probes with a time-to-live of 1, then 2, then 3, and so on. Each router decrements the TTL, and the one that drops it to zero discards the probe and sends back an ICMP "time exceeded" message from its own address. Line up those senders and you have the path, up to 30 hops with three probes each by default.
# Linux: UDP probes to high ports (the default), no DNS lookups
traceroute -n 203.0.113.10
# ICMP echo probes instead, like Windows uses
traceroute -n -I 203.0.113.10
# TCP SYN to port 443, which gets through firewalls that drop the others (needs root)
sudo traceroute -n -T -p 443 203.0.113.10
# Windows
tracert -d 203.0.113.10
The probe type matters more than it looks. A firewall that drops UDP to high ports turns the end of a default traceroute into a column of stars, while the TCP version to a port the firewall allows reaches the target. When a trace dies partway, try the other methods before you conclude anything.
Reading the output:
* * *means no reply came from that hop. Routers that do not send "time exceeded", or rate-limit it, look like this while forwarding traffic perfectly. It is not loss unless later hops also stop answering.- The three times on each line are three separate probes, each a round trip to that router.
- Two different addresses at one hop are normal where traffic is spread across equal-cost paths. Each probe can take a different branch.
- Linux traceroute sends up to 16 probes at once (
-N 16), which can trip rate limits and produce false stars.-N 1is slower and cleaner.
The deeper limit is direction. Traceroute shows only the forward path. The reply from hop 6 comes back by whatever route hop 6 prefers, which may be a different one, so a latency jump at hop 6 can live on the return path. Internet routing is often asymmetric, and the only way to see the other direction is to run the trace from the other end.
The mtr command: traceroute that keeps going
mtr discovers the path the same way and then keeps probing every hop, once per second by default. Interactive mode is useful at the desk; report mode is what you paste into a ticket:
# 100 cycles, wide hostnames, no DNS, then print a report and exit
mtr -r -w -n -c 100 203.0.113.10
# TCP probes to 443 when ICMP is filtered (may need root, depending on install)
mtr -r -w -n -T -P 443 -c 100 203.0.113.10
# UDP probes
mtr -r -w -n -u -c 100 203.0.113.10
-r is report mode, -w keeps long hostnames intact, -n skips DNS lookups (faster, and reverse DNS is often
wrong anyway), and -c sets the cycle count. The report default is 10 cycles, too few to measure anything below
10 percent loss, so use 100. -i changes the one-second interval; only root may go below one second.
A report from a monitoring host to a remote office looks like this:
Start: 2026-11-12T09:14:02+0000
HOST: mon-01 Loss% Snt Last Avg Best Wrst StDev
1.|-- 10.20.0.1 0.0% 100 0.4 0.5 0.3 1.9 0.2
2.|-- 198.51.100.1 0.0% 100 1.8 2.1 1.6 6.0 0.6
3.|-- 198.51.100.45 12.0% 100 14.9 22.7 11.8 98.4 17.9
4.|-- 203.0.113.1 0.0% 100 12.3 12.6 12.0 15.1 0.5
5.|-- 203.0.113.10 0.0% 100 12.6 12.9 12.2 16.0 0.6
Loss% is the share of probes that got no reply from that hop. Snt is probes sent. Last, Avg, Best and
Wrst are round-trip times in milliseconds, and StDev is their spread. This report is healthy. Hop 3 shows 12
percent loss and a messy average, but hops 4 and 5 show none, so every probe that passed through hop 3 arrived.
Hop 3 is simply slow to answer probes addressed to itself. Compare a report where the problem is real:
3.|-- 198.51.100.45 6.0% 100 15.1 15.6 14.9 21.0 0.9
4.|-- 203.0.113.1 7.0% 100 12.9 13.4 12.1 19.7 1.2
5.|-- 203.0.113.10 6.0% 100 13.0 13.5 12.4 20.2 1.3
Here the loss starts at hop 3 and carries through to the destination. Something on the link into hop 3, or on the way back from it, is dropping packets.
Reading an mtr report without fooling yourself
- Read the last line first. If the destination shows 0 percent loss and sane latency, the path is fine, whatever the middle rows say.
- Loss is real only when it starts at one hop and persists to the end.
- Loss or high latency at a single middle hop, gone by the next one, is that router's control plane, not your traffic.
- A latency step that persists (say 12 ms through hop 5, then 80 ms from hop 6 on) is either distance or a congested link. Light in fiber covers roughly 100 km per millisecond of round trip, so 70 ms between neighbors in one city is congestion or a detour, and 70 ms across an ocean is physics.
- Rising
StDevwith low loss usually means queuing: a link that is busy but not yet full. - If you can, run mtr from both ends. The provider will ask for both directions anyway.
On Windows, pathping gives a similar per-hop table without extra software. By default it sends 100 queries to
each hop, 250 ms apart, so expect about 25 seconds per hop before you see a result.
What none of the three can tell you
- Whether the application works. All three test packet delivery, nothing above it.
- How much bandwidth the path has. Use
iperf3between two hosts you control. - What happens inside tunnels and MPLS cores, which can hide several physical hops behind one line or none.
- What the return path looks like, except by tracing from the other side.
- What happened last night. These are live tools. For history, a long-running latency grapher such as SmokePing records loss and latency spread over weeks, which is often the evidence a provider needs.
A diagnostic order that holds up
ping -c 50the destination. No loss and normal latency: the network path is probably fine, look at the service.- Loss or high latency:
mtr -r -w -n -c 100from the affected side to the destination. - Trace stops early or the destination ignores ICMP: repeat with
-T -P 443, or a port the target serves. - Find the first hop where loss or latency starts and persists. That link, or the router after it, is the suspect.
- Run the same report from the far end if you have a host there, and attach both to the ticket with the time and your public address range.
- If everything is clean and users still complain, it is not the path. The troubleshooting flow picks up from there.
The one habit worth keeping: when you open a ticket about the network, attach a 100-cycle mtr report rather than a screenshot of a failed ping. It answers the provider's first three questions before they ask them. More pieces on the same theme live under network monitoring.
next: SNMP counter to bit rate for the numbers, the Rack list for the gear


