When a connection to another continent feels slow, the reflex is to blame bandwidth. Bandwidth is rarely the limit. Three other numbers decide what a link actually feels like: round-trip latency, packet loss and jitter. They show up as different symptoms, need different measurements, and only some of them can be fixed by choosing a better route.
Latency is a floor, not a fault
Round-trip time is the shortest time a packet needs to reach the far end and come back. It is set by physics plus the path: a detour through another region adds 50-150 ms before any congestion is involved. A link at 180 ms RTT will feel sluggish for keystrokes and remote desktop, no matter how much bandwidth you throw at it.
The useful question is not "how fast is it" but "how close is it to the floor". Measure the RTT to several nearby targets and to the actual destination. If the destination is 80 ms worse than a peer in the same city, the route is the problem, not the capacity.
Loss is what users call "stalling"
Packet loss below ~0.5% is mostly absorbed by TCP. Above 1% it stops being invisible: transfers stall, video buffers, SSH sessions freeze for a second and then catch up. Loss on a cross-border path is usually congestion at a peering point or a saturated transit link, and it is rarely symmetrical - the return path is often the worse one.
Loss is also the number most likely to be misread. A single ping timeout is not loss; you need a few hundred probes before the percentage means anything.
Jitter is what makes a call unusable
Jitter is the variation in latency between packets. Audio and video calls buffer a small window to hide it, so high jitter shows up as robotic voice and frozen frames even when average latency looks fine. A route with 120 ms average and 5 ms jitter is far more usable than one with 90 ms average and 60 ms of jitter.
# one pass, three numbers: loss per hop, average RTT, jitter
mtr -rwzbc 100 203.0.113.10
Read the report top to bottom. The first hop where loss stays constant and latency jumps is the segment you actually need to avoid.
What route selection can and cannot fix
Nothing fixes distance. If the destination is 9,000 km away, no client will bring the RTT under the physical minimum. What a routing layer can do is avoid the parts of the path that are congested or unstable:
- probe several entry points continuously instead of once at connect time;
- score each path on latency, loss and jitter, not on latency alone;
- move an established session to a better path without tearing the connection down;
- pin a route when a user asks for it, so a "randomly better" switch does not surprise them mid-call.
That last point is where most implementations get it wrong: a switch that resets the TCP connection is worse than the jitter it was trying to fix. The switch has to be invisible to the socket.
This is the part a 快连下载 client is responsible for - measuring the three numbers above on more than one path, and moving the session before the user notices. The measurement side is the easy half; doing the switch without dropping the connection is the hard one.
A short checklist
- Measure RTT against a nearby peer, not against a speed-test server.
- Use
mtrfor at least 100 probes before believing a loss number. - Treat jitter as its own metric - it predicts call quality better than average latency.
- When a path is bad, ask whether the client can switch session-preserving, or whether it needs a reconnect.
- Re-check after a fix: a route that improved at 10:00 can degrade again at 20:00 peak.
None of this needs new tooling. ping, mtr and a timestamped log of the three numbers will tell you more than any single "speed" figure, and it will tell you which of the three is actually costing your users time.
Top comments (0)