Why the fastest ping is often the slowest server
Ping measures the first hop. It says nothing about whether the gateway can still reach the internet, or how much it can carry once it does. What to measure instead.
Every server list sorts by ping, and ping is the weakest signal on the page.
A ping is a round trip to the gateway's front door. It tells you the machine is up and how far away it is. It does not tell you whether the machine can still reach the internet behind it, whether its upstream is saturated, or whether the route carrying your traffic is being shaped.
Three different numbers
- Latency is how long a round trip takes. It decides how a page feels while it loads, because a page load is dozens of round trips.
- Throughput is how much arrives per second. It decides how long a download takes, and it is nearly independent of latency.
- Loss is what fraction of packets never arrive. It quietly destroys throughput, because the usual congestion control reads loss as "slow down".
A gateway can answer in 20 ms and carry two megabits. Another can answer in 400 ms and carry two hundred. Sorted by ping, the list puts the wrong one first.
Loss is the one that hurts
On a long international path, a small amount of loss costs a great deal of speed. The default algorithm on most servers, CUBIC, treats every lost packet as a signal that the network is congested and backs off. On a path that drops packets for reasons other than congestion — shaping, a saturated exchange, a lossy last mile — it spends its life backing off from a problem that backing off does not fix.
BBR is the usual answer. It models the path's actual bandwidth and round trip instead of guessing from loss, and on a lossy long-haul link the difference is routinely several times the throughput. It is one setting on the server:
net.core.default_qdisc=fq
net.ipv4.tcp_congestion_control=bbr
That is worth doing before blaming the gateway.
What a client should measure
The honest measurement is a real request through the tunnel, timed. It costs a round trip and it is the only thing that reflects all three numbers at once: a gateway with a dead upstream fails it outright, and a gateway with a shaped route is visibly slower at it than its ping suggests.
Ranking on that, and keeping what was learned between sessions, gets you a list where the top entry is usually the right one. Ranking on ping gets you a list that is sorted, which is not the same thing.