You joined an Internet exchange (IX). The sessions are up. Now someone asks how you’ll know when something breaks, and your first instinct is to ping your peers (and/or the route servers).
It’s a good instinct with a bad target. Here’s how to do better.
Why pinging the IX fails
Start with the obvious problem. The peering LAN is supposed to stay on the fabric. Nobody should announce it into their own network, so your monitoring server usually can’t reach it. Some exchanges go further and block outside access on purpose, because a scannable peering subnet invites neighbor cache exhaustion and worse.
Some networks leak it anyway, which is why a few of you can ping those addresses today. That’s a mistake you’re borrowing from someone else. It also fills the exchange with junk. Every stray ARP or neighbor solicitation for an address that shouldn’t exist costs every router on the fabric a little bit of work, forever.
Now suppose the ping answers. You’ve asked a router about itself. That request goes to the control plane, which treats it as low priority. A slow reply may mean the router is busy. A fast reply means it’s awake. Neither tells you what your customers’ packets experienced crossing it, and the return path may not be the one you think you’re testing.
Think of it as rating a hotel by how fast the front desk picks up.
What to watch instead
Check session health where it lives, on your own router. BGP state to the route servers and to your peers, interface errors and utilization, optical levels. Poll them over your management interface with SNMP or streaming telemetry. LibreNMS does this well with zero licensing costs.
Measure what you reach through the exchange. Pick a handful of destinations you learn from peers there and probe those. They’re what your customers care about, and they’re what the path exists to deliver.
Chasing short drops
Intermittent trouble is the hard case. A customer reports that something stalls for a second or two, and you need to know whether the problem is inside your network or beyond it.
Smokeping (or Vaping) is a fine start, but it averages over time and can miss a brief event. For short drops, use mtr.
Run it from your side toward a destination you reach over the exchange. If the customer will cooperate, run it from their side too, because the two directions often disagree. Leave it looping in report mode in a screen session. When the problem shows up, take a screenshot.
A word of caution on probing every hop. Paths change, and a ping to a router along the way doesn’t behave like traffic passing through it. Probe the endpoints you care about, then use mtr to see what’s in between.
When you need more
You’re measuring from inside your own network, which is one vantage point. A drop that’s invisible from where you sit may be obvious from somewhere else.
RIPE Atlas gives you those other views. It’s a global network of volunteer-hosted probes, and you can run pings and traceroutes toward your destinations from probes in other networks. It’s free. Host a probe and you earn credits to spend on measurements. Run a few from probes near the destination or on the far side of the exchange, and you can start to tell whether the trouble follows your path or everyone’s.
It won’t catch a one-second stall. The sampling is coarse and it’s no substitute for your own monitoring. Treat it as a second opinion.
For deeper history and many vantage points at once, commercial tools exist that do this properly, and some offer free trials. If you’re a small network chasing one stubborn complaint, a trial may be enough to settle it.
Set expectations
You may not find the cause. Short drops are often elsewhere on the path, or gone by the time anyone looks. Good monitoring narrows the search. It doesn’t promise an answer. Tell your customer that early.
If you run an IX
Publish the rule so nobody has to guess: the peering LAN doesn’t leave the fabric. An AS0 ROA helps stop propagation but won’t catch leaks inside a member’s own network. Ingress ACLs on member edges will. Upgrade time is when you have the most leverage, so use it.
