DPI (Deep Packet Inspection) identifies which application each flow belongs to by looking beyond IP and port headers at protocol structure and the metadata exposed while a connection is set up. Service providers use DPI to see their traffic broken down by application, apply application policy at peak hours and plan capacity with real data. The purpose of modern DPI is classification; content is not stored.

What makes DPI different

Today most web, video, gaming and business applications share the same ports: TCP and UDP 443. A port number alone tells you very little about what the traffic is. Compare three approaches:

Port-based filteringFlow export (NetFlow, IPFIX)DPI
What it looks atL3/L4 headersSummaries of L3/L4 headersHeaders, protocol structure, handshake metadata, flow behaviour
Application recognitionWeakWeak; guesses by IPStrong
Real-time actionBlockingNone; reporting onlyPriority, shaping, rate limits, blocking
Processing loadLowLowHigher

Flow export answers "who talked to whom, and how much" cheaply and is ideal for DDoS detection. DPI answers "which application is this?" and can act on the answer.

Service providers use this visibility in three main places:

  • Capacity planning: see which applications grow at peak and time upstream investment accordingly.
  • Peering decisions: see which content networks your traffic goes to and decide whom to peer with at an IX.
  • Subscriber complaints: look at the traffic mix of a subscriber who says "the internet is slow" and tell whether the cause is the line, a device or an application.

How DPI classifies traffic

Protocol and signature recognition

Unencrypted protocols (DNS, HTTP, SIP and many P2P protocols) are recognised by their structure and signatures in the first packets. DPI also follows the expected message sequence of a protocol to cut false matches.

Metadata in encrypted traffic

Encrypted content is not visible, but some information is exposed while the connection is being set up. The SNI (the server name being contacted) and ALPN in the TLS ClientHello are the strongest clues for recognising an application. With QUIC, the initial packets are protected in a way that still lets network devices read the SNI. None of these methods decrypt anything.

Flow behaviour

Packet sizes, inter-packet timing, the ratio of traffic in each direction and the number of connections distinguish a video stream from a video call or a large download. As Encrypted Client Hello (ECH), which hides the SNI, spreads, behavioural analysis becomes more important.

Context

Correlating flows with DNS responses and using the known IP ranges of popular services improve classification accuracy.

From visibility to action: application policy

Once a flow is classified, policy can be applied to it:

  • Prioritisation: move latency-sensitive traffic such as VoIP, video calls and gaming to the front.
  • Shaping: cap the total bandwidth of an application group by queuing it.
  • Rate limits: upper limits per subscriber or per application.
  • Blocking: stop unwanted applications or categories.

Example rules:

  • Between 20:00 and 23:00, shape large downloads and P2P traffic while video and gaming flow normally.
  • Give VoIP and video conferencing a separate priority class for business subscribers.
  • Design application-aware plans, such as one that prioritises gaming traffic.

A good start is to collect visibility data for a few weeks before writing any policy. A rule written without knowing what fills the link at peak simply moves the problem elsewhere. Take local traffic-management and net-neutrality rules, and your subscriber contracts, into account as well.

Inline or mirror port?

InlineMirror port or TAP
VisibilityFullFull
Policy enforcementYesNo
Effect on trafficIn the traffic path; needs failover and bypassNone
RolloutPlanned cut-overQuick and risk-free

The practical path is to start in visibility mode from a mirror port and move inline once the need for policy is clear.

Watch two things when placing DPI:

  • See both directions: if asymmetric routing means only one direction of a flow passes the DPI, classification suffers.
  • Subscriber identity: per-subscriber reports require the DPI to know the subscriber. After NAT only the public IP and port are visible, so the data has to be correlated with BNG and CGNAT records.

Privacy and data retention

DPI inspects packets only to classify them. In a well-designed system content is never stored or reconstructed; only flow and application metadata are kept. Because even that metadata can be tied to subscribers it may count as personal data: restrict access to reports, define a retention period and review it against data-protection law such as GDPR or Türkiye's KVKK.

Sizing

DPI capacity is driven by inspected throughput, concurrent flows, new flows per second and the number of policies applied. DPI needs more processing than plain forwarding, so size it to your real peak-hour traffic profile.

With CoreGate DPI

CoreGate DPI performs L7 application classification, including on encrypted traffic, and applies policy, shaping and rate limits by subscriber plan, time window and application. It produces capacity reports by application, subscriber and time, and lets you export them to your own analytics. It runs inline or from a mirror port, and content is never stored. It comes with clusters of up to 16 nodes, automatic failover and the central management system, and is licensed by inspected traffic volume. For subscriber identity it can be deployed together with CoreGate BNG.