What Latency Should You Expect? Reference Baselines by Path Type
"Is 38 ms good?" It is the most natural question in networking, and it has no answer — not until you add three words: to reach what? A 38 ms round trip to a server on the next floor means something is badly wrong. The same 38 ms across a continent is excellent, and physically close to the best possible. Latency numbers only mean something when you compare them to what that kind of path should deliver. Most admins carry those expectations as folklore picked up from years of pinging things.
This guide is the folklore written down. It gives a reference table of typical round-trip times for every path type you are likely to meet. These range from two machines on the same switch to a link bounced off a satellite. For each one, it explains why the range is what it is and what a number outside the range usually suggests. Keep it bookmarked. Then it goes one step further, because a reference table is the beginning of judgment, not the end. You will learn to read a traceroute against these baselines, and to build your own network's "normal." You will also learn how much total delay real applications — a phone call, a remote desktop session — can actually tolerate.
In this guide
- How to read these numbers
- The reference table: typical RTT by path type
- Why the ranges are what they are
- The add-ons: last mile, Wi-Fi, and VPN
- Baselining: turning ranges into your normal
- Reading a traceroute against the baselines
- Latency budgets for real applications
- Where to go next
How to read these numbers
Four ground rules before the table, so it informs you instead of misleading you. First, every figure is a round-trip time (RTT) — out and back, the number a ping reports — measured on a reasonably quiet path. Second, these are ranges, not promises. Real paths vary with routing, equipment, and the hour of the day. A healthy link can live at either end of its range. Treat the ranges as "no cause for alarm" zones, not as service guarantees to wave at anyone. Third, your own measured normal beats this table, always. The table tells you what is plausible for a kind of path; only baselining (covered below) tells you what is normal for your path. Fourth, latency varies with load and time of day. The same link can measure 14 ms at dawn and 60 ms during the afternoon crunch. Both readings are honestly telling you about different traffic conditions, not contradicting each other.
The reference table: typical RTT by path type
| Path type | Typical RTT | Why | Outside the range suggests |
|---|---|---|---|
| Same switch | < 1 ms (often 0.2–0.5) | Meters of cable; switching in microseconds | Duplex mismatch, a struggling NIC or driver, an overloaded host answering slowly |
| Across the LAN (a few switches, one router) | 1–2 ms | Short distance; small per-device processing | A congested uplink, a router/firewall hop doing heavy work, a sick port |
| Across a campus | 1–5 ms | A kilometer or two; more hops | A saturated inter-building link; a loop or misrouted path |
| Metro / same city | 2–10 ms | Tens of km of fiber; provider equipment | Traffic taking a detour through a distant hub; congestion at the provider edge |
| Regional (a few hundred km) | 5–20 ms | Propagation starts to matter | Indirect routing; a congested peering point |
| Cross-country / continental | 30–80 ms | Thousands of km; ~1 ms RTT per 100 km of fiber | Below range: impossible, check the test. Far above: routing detour or congestion |
| Intercontinental | 70–250+ ms | Ocean cables; e.g. transatlantic ~70–90, transpacific ~100–180 | Above: hairpin routing through a third continent; satellite segment hiding in the path |
| Geostationary satellite | 500–700 ms | ~72,000 km up and down; ~478 ms is pure physics | Below ~480 ms: it is not geostationary. Far above: congestion on top of physics |
| Low-earth-orbit satellite service | ~25–60 ms, variable | Satellites only hundreds of km up, but moving; paths shift constantly | Sustained high or wild swings: obstruction, cell congestion, weather |
| Home broadband last mile (adds to any path) | +2–40 ms | Access technology overhead — see below | Large or growing additions: line trouble or an uplink saturated by the household |
| Wi-Fi hop (adds) | +1–10 ms, spiky | Shared airtime, retransmissions, power saving | Constant large additions or 100+ ms spikes: interference, congestion, weak signal |
| VPN (adds) | +1–5 ms, plus any detour | Encryption is cheap; the routing detour often is not | Tens of extra ms: traffic hairpinning through a distant concentrator |
A note on the two "impossible" cells: latency has a hard physical floor set by distance. So a reading below the floor for a path you believe is long means your belief is wrong. You are actually hitting something closer (a cache, a proxy, a local endpoint answering on the service's behalf) or your test is broken. That is not pedantry; noticing an impossibly good number has unmasked many a transparent proxy.
Why the ranges are what they are
Every row in the table is built from the same four ingredients: propagation, serialization, queuing, and processing delay. They get a full guide of their own in the four delay types. The short version, applied to the table: the floors of the ranges come from propagation, the travel time of light in fiber. Light in fiber covers only about 200 km per millisecond. Work one example to see it. A 4,000 km cross-country fiber route costs about 20 ms one way, 40 ms round trip, before a single device has added anything. Real routes wander along railways and seabeds, so they are longer than the map suggests. That is why the continental row starts at 30–40 ms and no vendor's brochure can start it lower.
The widths of the ranges come mostly from routing and queuing. Routing matters because two providers may carry the same city pair over very different cable paths. A "metro" packet that detours through a hub two hundred km away collects a regional-sized propagation bill. Queuing matters because delay collected waiting in device buffers is the one ingredient that changes minute to minute. It is why the same path reads differently at 6 AM and 3 PM. Serialization — the time to clock a packet's bits onto a link — only matters on slow links, which today means last-mile uplinks and legacy WAN circuits. Processing — device think-time — only matters when a box in the path is working hard (deep inspection, NAT on a weak CPU) or overloaded. When a measurement sits far above its row and it is not distance, it is nearly always queuing somewhere along the path. And if the inflation appears only when the link is busy, you may be meeting bufferbloat, the oversized-buffer disease that gets its own guide.
The satellite rows deserve a brief explanation each, because they are the table's physics lesson. A geostationary satellite must orbit at about 35,786 km so it can hover over one spot. Your packet climbs to it and descends twice per round trip, roughly 143,000 km of travel. Even at the full speed of light that is ~478 ms before any equipment gets involved. That explains the 500–700 ms row, and why nothing interactive ever feels good over geostationary. Low-earth-orbit constellations fly at a few hundred km instead, collapsing the propagation bill to LAN-of-the-sky levels. The price is constant motion, hand-offs between satellites, and therefore more variation than a buried fiber will ever show.
The add-ons: last mile, Wi-Fi, and VPN
The bottom three rows of the table are not paths of their own — they are surcharges added to whatever path lies beyond them. They matter because they sit at the very first hops, where every user's every packet pays them.
The last mile has a personality per technology
The link between a home or small site and its provider adds latency that depends on how the technology works, not just how many megabits it sells. Cable-modem systems schedule uploads — a modem must effectively request permission to transmit — which typically adds several milliseconds and a bit of jitter. DSL lines often run error-correction interleaving that trades latency for reliability, adding roughly 5–30 ms depending on settings and line quality. Passive optical fiber to the home is the quiet one, usually adding only a millisecond or two. Cellular access varies the most: scheduling, signal conditions, and distance to the provider's core can add tens of milliseconds and swing widely. None of these are faults; they are the character of the medium. The fault begins when the surcharge grows. The classic cause is the household or branch saturating its own uplink, queuing everything behind a cloud backup.
Wi-Fi is a conversation in a crowded room
Wireless adds a hop where devices share airtime and take turns. They retransmit whenever a frame is lost to noise. And — on battery-powered clients — they doze between packets to save power. The result: a healthy Wi-Fi hop adds a few milliseconds typically, but spikes are part of its nature. Occasional 50–100+ ms outliers on an otherwise fine wireless network are common where the same outliers on a cable would be alarming. The practical rule: whenever you are measuring latency to diagnose something, test from a wired machine too. That way, you know whether you are studying the network or studying the air.
VPNs: cheap crypto, expensive detours
A VPN (virtual private network — an encrypted tunnel between sites or from a laptop to the office) adds two costs. The encryption itself is small on modern hardware: a millisecond or few, more on an overloaded concentrator. The detour is the one that hurts: a full-tunnel VPN sends all traffic to the concentrator first. So a branch user reaching a nearby cloud service may travel branch → headquarters → cloud → headquarters → branch. If headquarters is 30 ms away, that hairpin adds 60 ms to every round trip. That is pure propagation tax, invisible in any device graph, obvious the moment you compare traceroutes with the tunnel up and down.
Baselining: turning ranges into your normal
The table answers "is this plausible?" Only a baseline — your own network's measured normal — answers "is this usual?" A baseline turns every future argument ("the WAN feels slow") into a comparison ("the WAN is at 46 ms; its normal for this hour is 18"). Building one is unglamorous and takes about fifteen minutes of setup:
- 1. Pick a handful of representative targets. A sensible starter set: your default gateway, your core switch, your main file or application server, the far end of each WAN or VPN link, and one steady external anchor beyond your ISP.
- 2. Measure on a schedule, not on a whim. A ping every minute to each target is plenty — gentle on the network, rich enough to see patterns. Ad-hoc tests only ever capture the moment you happened to be curious.
- 3. Record for at least two weeks. Long enough to see the daily rhythm, the weekly rhythm, and oddities like the nightly backup window.
- 4. Summarize honestly. For each target note the typical value and the spread — ideally the median and the 95th percentile, by time of day. (Averages hide exactly the spikes you care about; the measuring latency guide shows the percentile math and the tooling.)
- 5. Write the numbers down where future-you will find them. A one-page "normal card" per site — even in a wiki table — is the artifact. During an incident, nobody has time to derive normal from scratch.
Expect the baseline itself to have a shape, not a single value. Business networks breathe with the workday: quiet floors overnight, swells at 9 AM, lunch dips, backup-window humps. A number is only "high" against the right hour's normal. For example, 40 ms at 3 PM might be Tuesday as usual, while the same 40 ms at 6 AM might be your first clue that something changed overnight.
Tip: alert on deviation from your baseline, not on this table's ranges. A path that normally runs 8 ms and suddenly runs 25 ms is screaming at you even though 25 ms looks innocent on the reference table. The table's job is to sanity-check paths you have never measured; the baseline's job is to catch change. You need both.
Reading a traceroute against the baselines
Traceroute (the tracert command on Windows, traceroute on Linux and macOS) maps the routers along a path and shows the round-trip time to each one. Its real power appears when you read the per-hop numbers against the baseline table. The hop where latency steps up tells you what kind of segment your packets just crossed — and whether the step is geography or a problem. Here is a trace from an east-coast office to a European data center, lightly tidied:
1 10.0.0.1 (office gateway) 0.6 ms 0.5 ms 0.6 ms 2 203.0.113.1 (ISP first hop) 9.2 ms 8.8 ms 9.4 ms 3 203.0.113.65 (ISP metro core) 11.1 ms 10.9 ms 11.3 ms 4 198.51.100.14 (ISP backbone) 12.0 ms 12.4 ms 11.8 ms 5 198.51.100.77 (landing far side) 88.3 ms 88.9 ms 88.1 ms 6 192.0.2.33 (EU provider) 89.5 ms 90.2 ms 89.6 ms 7 192.0.2.161 (data center edge) 90.4 ms 90.1 ms 90.6 ms
Read it hop by hop with the table in hand. Hop 1 at 0.6 ms: LAN row, healthy. Hop 2 at ~9 ms: there is the last-mile surcharge, right where it belongs. Hops 3–4 around 11–12 ms: metro range, fine. Then the jump — hop 5 leaps from 12 to 88 ms, a +76 ms step. Alarming? Check the row: an ocean crossing at 70–90 ms RTT is exactly the transatlantic baseline. The jump is geography, not a fault; note that hops 6 and 7 confirm it by adding only small change on top. The finished trace reads as: a healthy path whose latency is where the map says it must be.
Three rules make traceroute reading honest. Rule one: only trust a step that persists. A real segment cost (like hop 5's ocean) carries forward into every later hop. If one middle hop shows 240 ms but the hops after it drop back to 90, that router is merely slow at answering traceroute. Routers treat "reply to probes about me" as the lowest-priority chore, handled by their busy management CPU. In that case, the router is forwarding your actual traffic at full speed. Ignore it. Rule two: judge each step against the path type it represents, using the reference table. Steps of +1 ms entering a campus, +10 crossing a metro area, and +80 crossing an ocean are all "correct." A step of +60 between two routers in the same city is a detour or a queue and worth a question. Rule three: remember the return path is invisible. Traceroute shows the outbound route, but every RTT includes a return trip that may travel a different way entirely. When a trace looks clean and latency is still high, ask someone at the far end to trace back toward you. (Traceroute has more caveats and more tricks than fit here — loss patterns in traces are covered with the packet loss guides.)
Latency budgets for real applications
Baselines describe what a path delivers; budgets describe what an application tolerates. The gap between them is your headroom, and knowing both turns "will voice work to the branch?" from a gamble into arithmetic. Plain guidance, end to end:
| Application | Comfortable (RTT) | Degrading | What breaking it feels like |
|---|---|---|---|
| Voice calls (VoIP) | < ~200 ms | ~200–300 ms | Talking over each other; walkie-talkie rhythm |
| Video meetings | < ~200 ms | ~200–400 ms | Awkward pauses; interruptions collide |
| Remote desktop / terminal sessions | < ~30 ms feels local; fine to ~100 | ~100–200 ms | Typing feels underwater; dragging windows lags |
| Interactive web apps | < ~50 ms | ~50–150 ms | Each click costs many round trips, so pages take seconds |
| Online gaming | < ~50 ms | ~50–100 ms | Actions register late; fairness complaints |
| File transfer, streaming, email | Very tolerant | — | Throughput sags on long paths, but buffers hide the delay itself |
Three footnotes that keep the table honest. Voice budgets are traditionally written as one-way, mouth to ear, with long-standing telephony guidance putting comfortable conversation under about 150 ms one-way. The RTT figures above are that guidance restated in the number you can actually measure. Latency is also not the whole story for real-time audio. Its variation (jitter) and packet loss can wreck a call that the latency budget says should be fine. And the web-app row hides a multiplier: one click can spend a dozen or more round trips on lookups, handshakes, and follow-up requests. That is why 100 ms of RTT can become a two-second page. For more, the round-trip anatomy guide dissects exactly where. Budget against your worst regular hour — the 95th percentile, not the average — because the average never annoyed anyone; the spikes did.
Where to go next
You now have the three layers of latency judgment. You have reference ranges for what any path type can deliver and a baseline for what your paths do deliver. And you have budgets for what your applications need. When a number worries you, the follow-ups are all here in the library. The guide to measuring latency honestly makes sure the number itself is trustworthy. The guide to the four delay types explains which ingredient of the number can be changed. And the high-RTT troubleshooting guide hunts down the cause when your baseline says something genuinely moved. The whole set lives on the Network Latency pillar page.
Put this into practice with SysPeep Monitor
A baseline only exists if something is measuring around the clock. Spot checks can't tell you what normal looks like at 3 PM on a Tuesday. SysPeep Monitor pings and polls your gateways, WAN links, and key servers on a schedule. It keeps the history and graphs each path's normal so deviations stand out on a dashboard. Alert rules with thresholds and dependencies mean you hear about the drift without drowning in noise. The free trial installs on a Windows machine in minutes.
Download Free TrialLearn About SysPeep