Home › Topics › Video Conferencing Performance › Service Reachability

Monitoring Reachability to Conferencing Services


The conferencing service lives in someone else's data center. You cannot log into its servers, watch its graphs, or reboot anything when it misbehaves. What you can do, and what your users actually experience, is reach it from your network, or fail to. Almost every "video is broken" morning comes down to one question. Is the path from us to the service healthy, or is the service having a bad day? You should be able to answer in seconds.


Answering that quickly is the whole job of reachability monitoring. It means running small automated checks from your own network toward the service, continuously. When trouble arrives, you already have a timeline instead of a stopwatch. The trick is knowing what to check. A conferencing service is not one thing at one address. It is name lookups, sign-in servers, and, the part most people miss, media servers that prefer a protocol many networks quietly break. I missed that part myself for longer than I would like to admit.


This guide covers what to check and in what order: DNS, then TCP and TLS, then the UDP media ports that matter most. It covers the firewall and NAT behaviors that damage media without blocking it. It explains how to measure latency to the service's media edge per site. It also covers using the provider's status page without trusting it blindly, and designing alerts that tell "our path" apart from "their outage."


In this guide

  • The service is a cloud you don't own
  • DNS: the first dependency
  • The synthetic checks, in order
  • Why UDP is the check that matters most
  • Firewall and NAT considerations
  • Measuring latency to the media edge
  • Correlating with the provider's status
  • The reachability check table
  • Alerts that separate our path from their outage
  • Where to go next

The service is a cloud you don't own


When you monitor a switch, you ask the switch how it is. You cannot do that with a service you do not own, so you flip the question around. Instead of asking the destination how it feels, you measure your journey to it. This is called synthetic monitoring. A small program pretends to be a user and performs the first steps of what a real client does. It records whether each step worked and how long it took.


A conferencing service is really several services wearing one name. There is a set of names to look up in DNS. There are sign-in and signaling servers, reached over TCP, the internet's reliable connection protocol. They are protected with TLS, the encryption layer. They handle logging in, scheduling, and the control channel that sets up a call. And there are media servers — the ones that carry the actual audio and video. They strongly prefer UDP. They are usually spread across regions so that each office talks to a nearby one. A reachability check that only pings a web address tells you almost nothing. The web front end can be perfectly healthy while the media path is broken. You have to check each layer, and you have to check the media layer the way media actually travels.


This guide is about reaching the service across the internet. If the service is unreachable because your own circuit is saturated, that is a capacity problem. Diagnose it with the outbound graph in What Video Calls Need from Your Network. It is not a reachability problem. Reachability monitoring assumes the circuit has room and asks whether the path itself works.


DNS: the first dependency


Nothing reaches the service until a name becomes an address. That makes DNS (the Domain Name System, the network's phone book) the first thing to check and the first thing to break. Conferencing services lean on DNS harder than most applications. They publish many names. They often return different addresses to different sites so each lands on a nearby media region. They change those answers frequently as they balance load. That is powerful and fragile at once.


Check DNS two ways. First, confirm your resolvers answer at all and answer quickly. A resolver that takes a full second to return a name adds that second to the start of every call. Second, resolve the service's actual names and sanity-check the answers. Check that they resolve and return addresses in a sensible region rather than a distant one. Check that they are not stuck on a stale answer after the service moved. A short scripted lookup shows the shape of it:


C:\> nslookup meet.example-service.com

Server:  resolver.corp.local
Address:  10.0.0.53

Non-authoritative answer:
Name:    meet.example-service.com
Addresses:  198.51.100.24
            198.51.100.25

Turn that into a monitored check: resolve the name every minute and alert if it fails or takes too long. Record the answer so that a sudden change of region is visible. Your London office might suddenly resolve to addresses an ocean away. When DNS is the problem, users say "everything is broken," and only a direct DNS check turns that into a ten-second diagnosis.


The synthetic checks, in order


Build the checks in the order a real client connects, because each layer depends on the one before it. A failure at a lower layer explains every failure above it. Run all of them from each site, on a schedule, all the time.


  • 1. DNS resolution. Resolve the service's names, as above. If this fails, nothing else can work, and the checks above it will fail for a reason that has nothing to do with them.
  • 2. TCP connect. Open a TCP connection to the sign-in and signaling endpoints on their ports (commonly the standard secure-web port). Success proves the path is open and the far end is listening. The time it takes is your round-trip to that endpoint.
  • 3. TLS handshake. Complete the encryption handshake and check the certificate is valid and not near expiry. A TCP connect that succeeds but a TLS handshake that fails points at inspection in the middle or a certificate problem, not a dead server.
  • 4. An application-level step, if you can. Fetch a health endpoint or complete a lightweight sign-in. This proves the service is not just accepting connections but actually answering, which a struggling service may not be.
  • 5. UDP media reachability. Confirm the UDP media ports are open and usable end to end. It is the check almost nobody sets up, and the one that predicts call quality best. Its own section is next.

Layering them this way means an alert points at the right layer. TCP fine but UDP blocked is a firewall problem on the media ports. DNS slow but everything else fine is a resolver problem. TLS failing alone is inspection or a certificate. The order turns a list of failures into a diagnosis.


Why UDP is the check that matters most


Real-time media travels over UDP (User Datagram Protocol), the internet's "send it and don't look back" delivery. Unlike TCP (Transmission Control Protocol), UDP just sends. TCP guarantees delivery by retransmitting anything lost and waiting for it to arrive in order. UDP sounds worse, and for a file it would be. But it is exactly right for a call: a video packet that arrives late is useless. There is no point retransmitting it and every point in not letting one lost packet hold up the rest.


Here is the trap. When a network blocks or mangles the UDP media ports, most clients do not fail. They quietly fall back to sending media over TCP, or through a relay, so the call still connects. Everyone congratulates the network for working. But media over TCP behaves the way TCP behaves. A single lost packet stalls everything behind it while it is retransmitted. On a real-time stream, that is felt as stutter, lag, and freezes. A call that has fallen back to TCP is a call that will be subtly bad all day. Nobody will know why, because "it connected."


This is why the UDP check is the most valuable one you can run and the most commonly missing. Confirm two things continuously. First, confirm that the UDP media ports are actually open outbound from each site to the service. Use a synthetic check that sends UDP to the service's media ports and confirms a response. Second, confirm that real calls are in fact using UDP rather than falling back. The service's per-call statistics almost always report the connection type per participant. A site showing "TCP" or "relayed" media has a firewall or NAT problem on the media path, no matter how green every other check looks.


Tip: the fastest way to discover a silent TCP fallback is to check any recent call's statistics for each site. Read the transport or connection type. If one office is on TCP or relayed while the others are on UDP, you have found a misconfigured firewall. You have found it before anyone has filed a single ticket — and fixed a problem the users had learned to live with.


Firewall and NAT considerations


The firewall is where media is most often quietly damaged, precisely because it usually does not block the call outright. Four behaviors matter; check and document all four per site.


  • UDP port ranges. Media servers use a range of UDP ports, published by the service. If the firewall only allows a subset, or allows none and forces the TCP fallback, quality suffers invisibly. Confirm the whole documented range is open outbound to the service, and re-confirm it after firewall changes, which is when it usually breaks.
  • NAT timeouts. NAT (Network Address Translation, which lets many inside devices share one public address) keeps a table mapping inside connections to outside ones. UDP has no "connection," so the firewall guesses when a flow is finished and removes its entry after an idle timeout. If that timeout is short, a quiet moment in a call can cause the mapping to expire mid-call. Someone might be muted, or there might be a pause in a screen share. The expired mapping drops media until the client reestablishes it. Long or generous UDP timeouts for the media ports prevent a whole category of mid-call drops.
  • Deep inspection that mangles media. Some firewalls try to inspect or "help" application traffic. On real-time media this ranges from useless to harmful: it adds latency. Features aimed at one protocol can rewrite or reorder media packets in ways that degrade the call. Where the service documents that inspection should be bypassed for its traffic, do so, and record that you did.
  • Session table capacity. A busy hour of meetings creates a large number of simultaneous flows through the firewall. If the session table fills, new flows fail and existing ones can be evicted. Monitor the firewall's session count against its limit, especially during meeting-heavy hours. That same session-table pressure is a general firewall concern worth watching year round.

The theme across all four is that the firewall can pass your TCP and TLS checks perfectly. Meanwhile, it can silently damage the media that rides UDP beside them. Its own logs will record this as a success. That is the reason the UDP check and the per-call transport type sit at the center of this guide.


Measuring latency to the media edge


Reachability is pass or fail; latency is how good the pass is. The service places media servers in many regions and steers each office to a nearby one. The meaningful measurement is the round-trip time from each site to the media edge that site actually uses, not to some central address. A branch office routed to a distant region will have calls that feel laggy even when nothing is "broken." Only a per-site latency measurement reveals it.


Measure it continuously from each site: the round-trip time to the media edge, and its jitter. Jitter is the variation in that time, which matters more for media than the raw number does. Keep the history so you can see three patterns. A steady, low, flat line is health. A line that is always high for one site suggests that site is being routed to a far region. That is worth a conversation with the service or a look at your own routing. A line that is usually flat but spikes at certain hours points back at your own congestion rather than the service. That is the capacity story in Giving Video the Road It Needs.


A check from your central monitoring server measures the path from headquarters. To know what a branch office experiences, the check has to run from the branch. A lightweight probe at each site is worth the modest effort. One probe per site turns "video is bad at the Denver office" from a mystery into a graph. The graph either indicts Denver's path or exonerates it.


Correlating with the provider's status


When several sites lose the service at once and your own path checks are green, the service is probably having an incident. The provider publishes a status page, and it is worth checking — but not worth trusting on its own. Status pages are often slow to turn red, scoped to regions in ways that may not include yours, and sometimes green during real trouble. Treat the status page as one witness, not the verdict. Status Page Skepticism explains why your own measurements outrank a vendor's dashboard when the two disagree.


At Tollgate Foods three sites lost media at 9:12 one morning. The provider's status page was green. So the network team spent forty minutes checking their own firewalls, resolvers, and circuits. All were fine, and they checked all of them twice. At 9:55 the page turned amber and announced a "degradation" that had begun at 9:10. The site probes had said so at 9:13.


The practical stance is to lead with your own evidence and use the status page to corroborate. Suppose your probes from three sites all fail to reach the media edge at the same minute, while DNS and your circuits are fine. You have an outage on their side whether or not their page admits it yet. You can tell your users so, with a timestamp, before the page updates. When your evidence and the page disagree, believe your measurements and note the discrepancy. You will be right often enough that the habit pays for itself.


The reachability check table


The full set in one place: what to check, how, how often, and what a failure of that check means. Run each from every site.


CheckHowFrequencyA failure means
Resolver healthResolve any name, measure the time1 minYour DNS is slow or down — everything will feel broken
Service name resolutionResolve the service's names; record the answers1 minDNS problem, or a stale or wrong-region answer
TCP connect to signalingOpen the port; measure the time1 minPath blocked, or the service's front end is down
TLS handshakeComplete the handshake; check the certificate5 minInspection in the middle, or a certificate problem
Application health stepFetch a health endpoint or light sign-in5 minThe service accepts connections but is not answering well
UDP media reachabilitySend UDP to the media ports; confirm a response1 minFirewall blocking media; calls will fall back to TCP
Media-edge latency and jitterRound-trip to the site's media edge1 minFar-region routing, or local congestion if it spikes by hour
Real-call transport typeRead connection type from per-call statisticsPer call / dailyA site on TCP or relayed media has a firewall or NAT fault
Firewall session countPoll the firewall over SNMP1–5 minSession table filling; new flows will fail in busy hours

Alerts that separate our path from their outage


A reachability system that cannot tell "us" from "them" produces alerts nobody can act on. The design goal is that the alert itself says whose problem it is. That means comparing sites and layers against each other rather than reading any single check alone.


  • One site fails, others fine. The problem is that site's path: its circuit, its firewall, its resolver, its route to the service. Alert the people who own that site.
  • All sites fail the same check at the same minute. Either a shared dependency you own (a central firewall, a shared resolver) or the service itself. If your shared components are healthy, call it a provider incident and communicate it as one.
  • UDP fails but TCP and TLS pass. A firewall or NAT problem on the media ports, almost always yours. This is a high-value alert precisely because calls will keep connecting and quietly underperforming.
  • DNS fails or slows. Alert distinctly, because it masquerades as a total outage and misdirects everyone who does not have the specific check.
  • Latency to the media edge rises for one site by hour. Route it to capacity planning, not to an outage channel; it is congestion, not breakage.
  • Transport type turns to TCP or relayed for a site. A quiet alert to whoever owns firewalls, since it indicates a silent fallback rather than a live outage.

Two refinements keep these trustworthy. Require a check to fail from more than one probe, or for more than one interval, before it pages. That way, a single blip does not page anyone. And model the obvious dependencies: every check depends on the site's circuit and resolver. When the circuit goes down, you then get one alert about the circuit, not a cascade of a dozen reachability failures underneath it.


Tip: the single most useful alert you can build here is "UDP media path down, everything else up." It fires before any user notices, and the problem is almost always your firewall. Fixing it removes the invisible tax that silent TCP fallback puts on every call from that site. If you build only one check from this guide, build that one.


Where to go next


You cannot see inside the conferencing service, but you can measure your journey to it, and that journey is what your users live. Check the layers in order (DNS, TCP, TLS, an application step, and above all the UDP media path) from every site. Watch the latency to each site's media edge. Read the transport type from real calls to catch silent fallbacks. Design alerts that compare sites so the alert names the culprit. Do that and "is it us or them?" has an evidence-backed answer in seconds. The status page will agree with you eventually. The rest of Video Conferencing Performance covers the other half: what to do when the path is fine but the circuit is full. It also covers how to give video the capacity it needs at the busy hour.


Put this into practice with SysPeep Monitor

Synthetic checks toward a service, from more than one site, are exactly what SysPeep Monitor is for. It runs reachability and health checks over the collectors it supports and keeps the history. It can run lightweight probes at branch sites. That way, each office's path to the service is measured from where the users are. Dependency rules keep a dead circuit from burying the reachability alerts underneath it. The free trial installs on one Windows machine.

Download Free TrialLearn About SysPeep