ntopng: Watching Where Your Bandwidth Actually Goes
A router graph tells you the pipe is full. It never tells you which device filled it.

Contents
My router’s traffic graph has told me the same thing for years: a big spike most evenings, a flatter hump during the day, and a WAN link that’s occasionally pinned at capacity for reasons it refuses to specify. It’s a graph of bytes in and bytes out. It has never once told me which of the eighteen devices on my network was responsible, and it never will, because a router counting interface octets doesn’t know what a “device” is, let alone what an “application” is.
That’s the gap Netdata and a Prometheus and Grafana stack don’t close either, because they’re both built around scraping metrics endpoints on hosts you control. They’re excellent at “how much CPU is this container using” and useless at “why did the guest VLAN do 40GB overnight.” Answering that requires looking at the packets themselves, on the wire, before they’ve been reduced to a counter. That’s what ntopng does, and it’s a genuinely different tool from a metrics stack rather than a competitor to one.
Why port numbers stopped being a reliable answer
The obvious way to classify traffic is by port number: 443 is HTTPS, 22 is SSH, 25 is SMTP. That worked reasonably well around 2005. It works badly now, for two reasons. First, almost everything has consolidated onto 443 — video calls, backups, game clients, telemetry, and actual web browsing all look identical to a tool that only reads port numbers. Second, QUIC and other UDP-based protocols have made even the transport-layer signature less informative than it used to be.

ntopng’s answer is nDPI, ntop’s own deep packet inspection library, bundled in and doing the classification work under the hood. Instead of trusting the port, nDPI fingerprints the actual byte patterns and TLS handshake metadata (SNI, JA3-style fingerprints, certificate fields) to identify the application behind a flow even when it’s tunnelled over 443. It’s not infallible — a sufficiently determined VPN or a service using domain fronting can still slip past classification into a generic “TLS” bucket — but for the ordinary traffic on a home or small-office network, it correctly identifies the flow eating 200Mbps as a cloud backup client, something the port number alone would only ever call “some HTTPS.”
Getting it traffic it can actually see
This is the part that trips people up, because ntopng doesn’t route or proxy anything — it’s a passive observer, and a passive observer can only see what’s put in front of it. On a normal switched network, a host’s network interface only receives its own unicast traffic plus broadcast and multicast, which is not enough to build a picture of the whole LAN. You need one of three things: a managed switch with a SPAN or mirror port that copies all traffic to ntopng’s interface, a network tap sitting inline, or ntopng running directly on the router/gateway box where all traffic already passes through one interface by construction.
The gateway option is the cheapest to set up and the one I’d point most homelabbers at first, because it needs no extra hardware — just capability on the interface:
| |
And the config file it reads, pointed at the bridge interface that actually carries LAN-to-WAN traffic on a router box:
| |
network_mode: host and the two added capabilities are non-negotiable — ntopng needs to put the interface into promiscuous mode and read raw packets, and Docker’s default bridge networking hides the real interfaces from the container entirely. If you’re mirroring from a managed switch instead of running on the gateway, swap -i=br-lan for whatever interface the mirror port lands on, and expect that interface to show nothing but mirrored traffic — it won’t have an IP of its own and doesn’t need one.
Reading the dashboard without drowning in it
The default view is a live table of hosts sorted by throughput, and it’s tempting to stop there because it already answers the “who” half of the question. The more useful screen is the Flows view, which breaks each host down into individual connections tagged with the nDPI-identified application, destination, and volume. That’s where “why is the office laptop doing 300Mbps at 2am” turns into “it’s a scheduled backup job talking to a cloud storage endpoint,” which is a completely different problem (and non-problem) than “it’s been compromised.”
Host Pools let you group devices — IoT, guest VLAN, work laptops — and get aggregate figures per group instead of per device, which matters once you’re past a handful of hosts and the flat list stops being scannable. I’ve found the pools most useful for exactly the boundary that egress firewalling is meant to enforce: a pool for “things that should only ever talk to a short list of destinations” makes it obvious within seconds when one of them doesn’t.
Keeping history without building a second database
ntopng keeps its own lightweight historical time series on disk by default, which is fine for the “what happened in the last few days” question but gets thin fast if you want months of trend data or the kind of dashboarding a proper time-series backend gives you. The community edition can export flow and alert data to a ClickHouse instance for longer retention and SQL-style querying — the same trade a lot of self-hosters make when ELK-style log tools turn out to be overkill for the actual volume they’re dealing with. If you already run ClickHouse for something else, pointing ntopng at it is a small config change; if you don’t, the built-in retention is genuinely enough to answer “who filled the pipe last night,” which is the question that actually gets asked.
When you still reach for tcpdump instead
ntopng answers “who and what,” aggregated and classified, and that’s a different job from “what exactly is in this one connection.” When a flow looks suspicious — a host talking to an address nDPI can’t classify, at a volume that doesn’t match anything scheduled — the next step isn’t more dashboard, it’s a packet capture on that specific conversation. I still reach for tcpdump and nsenter at that point, because ntopng will only tell you that 40GB went to an unclassified TLS endpoint overnight, leaving what was actually said in the handshake and whether the certificate looks legitimate for the capture to answer. Treat ntopng as the triage layer that tells you where to point the scalpel, leaving the cutting itself to a dedicated capture tool.
The two tools are complementary in the way a smoke detector and a fire investigation are complementary — one runs continuously and tells you something’s wrong right now, the other gets pulled out once you already know where to look. Running both is standard practice on my network: ntopng always on, watching every host, and a targeted capture only when a flow earns the extra attention.
Alerts that don’t need a second monitoring stack
Past a certain point, staring at a dashboard stops scaling, and ntopng’s community edition includes a basic alerting engine that saves the trip to Grafana’s Alertmanager for the things it can see that a metrics scraper can’t. You can set thresholds per host or pool — total throughput over a period, a flow to a country your household has no business talking to, a device suddenly appearing on a VLAN it’s never used — and get a flagged alert in the UI, or a webhook out to whatever actually pages you.
I keep this deliberately narrow: one rule for “any host exceeds 500Mbps sustained for five minutes” (which has caught exactly two real incidents in over a year, both misbehaving software updates rather than anything sinister) and one for “a device on the IoT pool talks to anything outside a small allow-list of destinations.” Anything more elaborate than that duplicates work Grafana’s alerting or Alertmanager already do better on the metrics you’re already collecting elsewhere, and running two alerting systems that can both page you is a good way to start ignoring both.
A real evening, worked through
The scenario that made me actually install this: a WAN graph pinned near saturation every night between 9 and 11pm, no obvious culprit, and a household of people all denying it was them. Netdata told me the box pushing the traffic was a mini PC in the living room. That’s where a port-based view stops, because the mini PC runs three different things and all of it looks like “outbound HTTPS” to anything that isn’t inspecting the payload.
ntopng’s flow view broke it down in under a minute: one flow, classified by nDPI as a cloud photo-backup client, sustained at just under 300Mbps for two hours straight, timed to start the moment the phone it was backing up came home and joined Wi-Fi. Nothing compromised, nothing malicious — just an upload schedule nobody had told the network about, and a setting in the backup app that was happy to run on a bandwidth cap once I found it. That’s a five-minute fix once you can see it, and an unsolvable mystery when all you have is a total-bytes graph.
Troubleshooting
The interface shows nothing. Almost always a capabilities or networking-mode problem rather than anything wrong in ntopng.conf itself. Confirm the container actually has NET_ADMIN and NET_RAW, and that it’s on host networking rather than Docker’s bridge — a bridge network gives the container a virtual interface with none of the host’s real traffic on it, so ntopng will report a live, empty interface and no errors.
Everything gets classified as generic TLS. Expected for traffic ntopng hasn’t seen a fingerprint for yet, and increasingly common as more services roll out encrypted client hello (ECH), which hides even the SNI field nDPI used to rely on. There’s no fix for this beyond accepting a “unclassified encrypted” bucket that will only grow over time; it isn’t a misconfiguration.
CPU usage climbs steadily on a busy link. nDPI inspection isn’t free, and on a gateway pushing several hundred Mbps of mixed traffic through a low-power box, the classification overhead is measurable. Dropping to the lighter fast-path classification mode, or moving capture to a dedicated low-power box fed by a mirror port instead of doing it inline on the router, both help.
A device I know exists never shows up. Check whether it’s on a VLAN that isn’t included in the mirror or bridge ntopng is watching — a common trap on a multi-VLAN setup where the router has several bridge interfaces and only one is configured in ntopng.conf.
Is it worth running
If your monitoring stack already tells you total throughput and per-container resource use, and what you actually want is per-host, per-application attribution — “which device, doing what, is using the bandwidth” — ntopng answers a question nothing else in a typical homelab stack does. It answers a different question than Prometheus and Grafana or a log aggregator do: a purpose-built lens on the wire itself, and the fastest way I’ve found to turn “the link is full” into an actual answer instead of a guess.




