Everything Ping Actually Tells You — and Five Things It Doesn't
Open a terminal on any machine on earth, and the odds are good the first diagnostic command ever typed on it was ping. It is the hammer of networking: always within reach, older than the web, and rarely ever explained. That gap matters, because ping gets quoted as evidence constantly. People say "the server pings, so it's fine," or "ping looks bad, the network is broken." Both sentences are wrong often enough to burn real troubleshooting hours.
This guide unpacks ping in two halves. First comes everything a humble echo really does tell you when you squeeze it properly. That includes reachability, distance hints, jitter clues, even the exact size of the largest packet a path can carry. Then come the five things ping cannot tell you no matter how hard you stare. Each comes with the kind of story that ends in an apologetic ticket update. The ping built into Windows and Linux (macOS behaves like Linux here) is the whole toolkit.
In this guide
- What a ping actually is
- Round-trip time: min, avg, max, and the underrated stddev
- The TTL column: distance hints and folklore
- Packet-size games: finding a path's real MTU
- Counts, intervals, and flood etiquette
- Reading loss patterns
- Five things ping does not tell you
- The flags worth knowing
- The doorbell, not the interview
What a ping actually is
Ping sends an ICMP echo request — "repeat this back to me" — and waits for the echo reply. ICMP, the Internet Control Message Protocol, is the network's built-in service channel: short control messages that IP devices use to report problems and answer basic questions. It rides directly on IP, alongside the TCP and UDP that carry your application traffic, not inside them.
One detail quietly explains half of this article: the reply comes from the target's operating system network stack. No application is consulted — the kernel hears the knock and answers it, whether or not anything useful is running upstairs. Keep that in your pocket.
When a reply comes back, you have proven four things at once, all at the IP layer — the layer that moves packets between addresses:
- Your own stack and interface are working.
- Some outbound path delivered one small packet.
- The target's stack is alive and willing to answer.
- Some return path delivered the answer.
Useful — but notice the qualifiers: one small packet, at that instant, answered by the stack, over some path. Each is a door into part two. First, squeeze the reply for everything it carries — a typical Linux run:
$ ping -c 10 fileserver01 PING fileserver01 (192.168.10.21) 56(84) bytes of data. 64 bytes from 192.168.10.21: icmp_seq=1 ttl=64 time=0.482 ms 64 bytes from 192.168.10.21: icmp_seq=2 ttl=64 time=0.391 ms ... --- fileserver01 ping statistics --- 10 packets transmitted, 10 received, 0% packet loss, time 9188ms rtt min/avg/max/mdev = 0.371/0.442/0.612/0.071 ms
Each reply line carries three facts. The field icmp_seq is the probe's sequence number, making gaps and reordering visible at a glance. We will decode the field ttl shortly. The field time is the round-trip time (RTT), the full out-and-back journey. Windows prints the same facts with different labels. The summary block is where reading begins.
Round-trip time: min, avg, max, and the underrated stddev
Most people read one number from a ping test: the average. Experienced people read four, because each answers a different question.
min is the floor — the trip when nothing got in the way. With no queues to sit in, travel time is mostly physical distance, so the minimum is your cleanest measure of how far away something effectively is. If yesterday's min to a branch office was 12 ms and today's is 48 ms, packets are taking a different road.
avg is the headline number, and like most headlines it hides the story: calm moments and congested ones blended into one bland figure. Fine for scale; useless for diagnosis alone.
max is the single worst moment — usually one probe that sat in a queue while a burst of other traffic drained ahead of it. One high max in hundreds of probes is life; a max routinely ten times the min means regular traffic jams.
mdev (Linux calls it mean deviation; macOS prints stddev) measures how spread out the samples were — the most underrated number in the block. Spread in delay is called jitter, and this field is your poor-man's jitter needle. Consider two links that both average 30 ms:
rtt min/avg/max/mdev = 29.1/30.2/31.4/0.6 ms <- calm and steady rtt min/avg/max/mdev = 12.3/30.8/96.5/22.9 ms <- same average, chaos
The first link is a metronome; the second swings wildly between 12 and 96 ms. A file download barely notices. A voice call, video meeting, or remote desktop session absolutely does, because real-time traffic cares more about consistency of delay than the amount. The stddev is how you check whether "fine, it averages 30" is actually fine.
One honest limitation: Windows prints only minimum, maximum, and average — no spread figure. If jitter is the question, reach for MTR, which reports a standard deviation per hop.
The TTL column: distance hints and folklore
TTL stands for Time To Live, and despite the name it counts hops, not seconds. Every IP packet starts with a TTL number, and every router that forwards it subtracts one. At zero, the packet is discarded. That is a self-destruct that stops lost packets circling the internet forever. (That self-destruct, cleverly abused, is the entire engine behind traceroute — the story of the next guide in this pillar.)
The ttl= in a ping reply is what remained after the reply crossed the network back to you. Operating systems start packets at a few conventional values. Most Unix-like systems start at 64, Windows at 128, and much network gear at 255. So a little subtraction gives a rough hop distance. A reply with ttl=118 most likely started at 128 and crossed ten routers. A reply with ttl=64 came from your own segment, no routers in between.
This shades into folklore: guessing the far end's operating system from its TTL family. Replies in the 50s "smell like Linux," around 110–120 "smell like Windows," 250-ish "smells like a router." Treat it as exactly that — folklore, since initial TTLs are configurable. But it earns its keep in one case. When you ping a "server" and the reply carries a router-flavored TTL, ask whether a firewall or load balancer is answering on its behalf. And if the TTL changes between replies in one test, different devices are taking turns answering, or the path is flapping. A light clue, but a real one.
Packet-size games: finding a path's real MTU
Ping's most underused trick solves a famously maddening class of problem. Every link has an MTU — Maximum Transmission Unit — the largest packet it will carry: 1,500 bytes on standard Ethernet. A path is a chain of links, and its effective MTU is the smallest in the chain. VPN tunnels and some broadband connections wrap your packets in extra headers of their own, shrinking the room left below 1,500.
A packet too big for a link meets one of two fates. The router chops it into fragments. Or — if it carries the DF (Don't Fragment) flag — the router drops it and sends back an ICMP complaint. That complaint says: "too big, and you told me not to cut it." Modern systems set DF on nearly everything and rely on those complaints to learn each path's limit (Path MTU Discovery). When an overzealous firewall blocks the complaints, discovery breaks and you get a black hole: small packets glide through, large packets silently vanish.
The symptom is unmistakable once seen. The VPN connects, ping works, and you can log in and type at prompts. But web pages half-load and file copies hang at 0%. Everything interactive is small packets; everything bulky rides full-size packets, which are falling into the hole. Ping's default payload is tiny (56 bytes on Linux, 32 on Windows), which is why "it pings" proves nothing here. With size and DF turned up, it becomes the perfect probe:
# Windows: -f sets Don't Fragment, -l sets payload size in bytes ping -f -l 1472 fileserver01 # Linux: -M do prohibits fragmentation, -s sets payload size ping -M do -s 1472 fileserver01
Why 1472? Payload plus 8 bytes of ICMP header plus 20 bytes of IP header is 28 bytes of overhead. So 1,472 makes a 1,500-byte packet — exactly full-size for plain Ethernet. If the whole path carries 1,500, this ping succeeds. If some link is smaller, you will see one of these:
Packet needs to be fragmented but DF set. (Windows) From 10.8.0.1 icmp_seq=1 Frag needed and DF set (mtu = 1400) (Linux, a router reports)
Best case, a router announces the answer, like the mtu = 1400 above. If the complaints are filtered, you get plain timeouts at large sizes while small sizes pass. That is not a dead end but a diagnosis: it is the black-hole signature. Either way, find the true limit by bisection — the higher-lower guessing game:
- 1472 fails. 1200 works. The limit is between them.
- Split the difference: 1336 works. 1404 fails. Narrowing.
- 1370 works, 1387 fails… 1372 works, 1373 fails: 1372 is the ceiling.
- Add back the 28 header bytes: path MTU 1,400 — a very tunnel-shaped number.
Five minutes of pinging just measured, to the byte, a property of a path crossing networks you do not control. The usual fix — telling the tunnel to advertise the smaller limit — belongs to your VPN's documentation, but you arrive holding the number it needs.
Warning: the flags collide across operating systems. On Windows, -f means "don't fragment." On Linux, -f means flood — send probes as fast as possible. On Linux, -l means "preload this many packets," not payload size. Pasting the Windows MTU test into a Linux shell therefore asks for a flood; thankfully Linux demands root for that. Learn the pairs, not the letters — the reference table near the end of this guide is the safety card.
Counts, intervals, and flood etiquette
The defaults differ: Linux ping runs until you press Ctrl+C, then prints its summary; Windows sends exactly four probes and stops. Four probes is a coin flip, not a measurement. So for anything you intend to quote, run a fixed, decent-sized count. Use ping -c 100 host on Linux, ping -n 100 host on Windows. Fixed counts make honest before-and-after comparisons.
Pace matters too. Linux sends one probe per second by default and speeds up with -i — -i 0.2 is five per second; anything truly rapid is reserved for root. Windows has no interval flag and plods at about one per second. One probe per second is background noise nobody notices. Dozens per second aimed at equipment you do not own looks like an attack, and intrusion-detection systems will happily agree.
Then there is the sledgehammer: sudo ping -f, flood mode. It sends probes back-to-back, at least a hundred per second, printing a dot per request and erasing one per reply. The dots left on screen are a live picture of loss. It is a legitimate stress test for your own lab gear in a maintenance window, and nothing else. Restraint also wins for a practical reason. Many devices answer ICMP lazily and rate-limit it when pressed. So the harder you hammer, the worse your numbers look. Your own aggression manufactures the "problem" you are hunting.
Reading loss patterns
"5% packet loss" is a summary. The shape of the loss says more than the percentage, and icmp_seq lets you see it. Run a few hundred probes and watch which sequence numbers go missing:
64 bytes from 203.0.113.9: icmp_seq=41 ttl=54 time=34.1 ms 64 bytes from 203.0.113.9: icmp_seq=42 ttl=54 time=33.8 ms 64 bytes from 203.0.113.9: icmp_seq=45 ttl=54 time=35.0 ms <- 43 and 44 are gone: a burst of two
Three shapes cover most of what you will meet:
- Metronome loss — every Nth probe missing with eerie regularity — suggests something deterministic. That could be packets spread across a bundle of parallel links with one sick member, or a rate limiter trimming on a rigid budget.
- Burst loss — runs of consecutive losses, clean stretches between — suggests time-correlated events. Those could be a queue overflowing during traffic spikes, pulsing Wi-Fi interference, a flapping link, or a network reconverging after a topology change.
- Sparse random singles — lone losses without rhythm — is the signature of physical-layer errors: a failing cable, a dirty fiber connector, dying optics, a marginal wireless signal.
Hold these as suggestions, not verdicts — the pattern says where to look, and the suspect device's interface error counters say whether you looked right. And one caveat looms: probe loss and traffic loss are not automatically the same thing, for reasons we are about to meet.
Five things ping does not tell you
Everything above was squeezed from a tiny echo. Now for the boundary: none of these five are edge cases; each is a daily mistake made by trusting ping one step past what it measures.
1. It does not tell you a service works
The 2 a.m. classic. The web server answers every ping in half a millisecond, and the ping-based check glows green. And the website has been down for an hour. Remember who answers echoes: the kernel. A machine whose application has deadlocked or whose disk is full will cheerfully ping all night. The reverse is just as common: a healthy server whose firewall drops ICMP, panicking a ping-only monitor while users work undisturbed. Ping proves the doorbell rings — not that anyone inside is doing their job. The full argument, and what to check instead, lives in ping up vs. service up.
2. It does not tell you the path your application takes
Ping's packets are ICMP. Your application's packets are TCP or UDP, aimed at a port — and networks are entirely capable of treating those differently. Policy-based routing can steer web traffic out a different link. Proxies and inspection appliances intercept application protocols but never see your echoes. VPN split tunnels send some destinations through the tunnel and some around it. So an administrator pings a cloud service, sees a pristine 18 ms, and declares the network healthy. Meanwhile users' actual HTTPS sessions are hairpinned through an overloaded inspection proxy adding 400 ms of misery. Both observations are true — measured on different roads. When ping and users disagree, believe the users and ask which path their traffic really takes.
3. It does not measure low loss rates accurately
Each probe is one sample, and percentages built on few samples are fiction. A hundred pings resolves loss only in 1% steps — and even that overstates it. Suppose a link truly loses 0.5% of packets, enough to drag down TCP throughput and put crackles in voice calls. The chance a 100-probe test loses nothing at all is 0.995 raised to the 100th power — about 61%. Six runs in ten hand you a spotless "0% loss" over a genuinely unwell link. Resolving fractions of a percent honestly takes thousands of samples — continuous measurement, not a coffee-break test. The sampling math and tooling live in measuring packet loss.
4. It does not tell you anything about one direction
Ping's time is a round trip: outbound plus return, fused into one number. A 90 ms RTT might be 45 out and 45 back — or 10 out and 80 back through a congested return route. Internet asymmetry is normal: the reply often travels a different road than the request, because each network chooses its own handoffs. This is the asymmetric-path trap: the sick direction hides inside a healthy-looking round trip. It bites hardest on real-time traffic, whose budgets are per direction. A call can be crisp one way and garbled the other while ping reports one bland average. Splitting the directions takes synchronized clocks and purpose-built tools; what ping honestly gives you is the sum, so treat it as the sum.
5. It does not promise your probe was treated like real traffic
Answering pings is a courtesy, and network devices treat it like one. Routers forward customer traffic in specialized hardware at wire speed. But ICMP aimed at the router itself is handled by its modest management CPU — low priority, rate-limited, answered when convenient. Some networks also mark ICMP down in their quality-of-service schemes — cheap seats end to end. The result: ping can report loss and latency that real traffic never experiences — a router "dropping 30% of pings" while forwarding flawlessly — and occasionally the reverse. That deceit, and how to design honest checks around it, is the whole subject of ICMP politics.
The flags worth knowing
Everything this guide used, in one place. The same intentions wear different letters on each system — four of them with entirely different meanings — so treat this table as a safety card:
| What you want | Linux | Windows |
|---|---|---|
| Send exactly N probes | -c 100 | -n 100 |
| Run until stopped | default (Ctrl+C to stop) | -t |
| Payload size in bytes | -s 1472 | -l 1472 |
| Forbid fragmentation (DF) | -M do | -f |
| Seconds between probes | -i 0.5 | not settable (about 1 s) |
| Wait longer for each reply | -W 5 (seconds) | -w 5000 (milliseconds) |
| Set the outgoing TTL | -t 10 | -i 10 |
The doorbell, not the interview
Here is the mental model that keeps ping in its proper place. Ping rings the doorbell. If someone answers, you know the address is real, the door works, and how long the walk to the door took. What the doorbell can never tell you is whether the kitchen is on fire, which route your guests will actually drive, or whether the residents answer parcels as promptly as doorbells. It is the first question of an investigation, never the verdict.
Used that way, ping is superb. It is a two-second test that sorts "unreachable at the IP layer" from "reachable, keep digging." It is also a pocket toolkit for distance, jitter, MTU, and loss-shape clues. The digging belongs to its relatives: traceroute shows which doors the packet knocked on along the way, and MTR turns anecdotes into per-hop statistics. The rest of the Ping, Traceroute & Friends pillar covers the whole family — including the ways the internet lies to all of them.
Put this into practice with SysPeep Monitor
Squeezing one ping is a skill; remembering to squeeze it every minute, forever, is a job for software. SysPeep Monitor keeps continuous watch on the network behind these tests. It auto-discovers your devices and collects health and performance metrics over SNMP, WMI, and SSH. It uses thresholds and dependency-aware alert rules to tell you what actually broke instead of paging you for everything at once. The free trial installs on a single Windows machine.
Download Free TrialLearn About SysPeep