Part 2 of 8Firewall Ecosystem
,

Physical Firewalls and Next-Generation Firewalls

A technical tour of what happens inside a firewall appliance, from first-packet inspection to offload, encrypted traffic and failover.

Read in: English · తెలుగు · हिन्दी · Français · Español

Sign in to save

A rackmount firewall appliance with vendor logos on its bezel sits below a glass tube of inspection stages that ends in a verified shield.

A physical firewall is more than a network port with allow-and-deny rules. Its processing architecture, inspection engines, acceleration paths and high-availability design determine which security services it can sustain under real traffic loads. This article explains what next-generation firewalls add and how to evaluate those capabilities beyond the headline throughput figure.

Physical firewalls and NGFWs: what are we examining?

A physical firewall is a dedicated network security appliance with processor resources, network interfaces, storage, firmware and management facilities designed to enforce traffic policy. A next-generation firewall (NGFW) is defined by its inspection and security functions, such as application identification, intrusion prevention and policy controls informed by more than packet headers. The terms describe different dimensions. An NGFW can be delivered as a physical appliance or a virtual service, while not every dedicated firewall offers the same NGFW capabilities.

Article 1, Firewall Fundamentals and Architecture explained packet filtering, connection tracking, firewall policies and NAT. Here the focus shifts from those foundational mechanisms to the engineering of the appliance itself. The essential questions are which processing resources handle each flow, how the security services change that path, what the device can inspect in encrypted traffic, and what happens when a node fails.

A firewall is an enforcement point, not a guarantee that every permitted connection is safe. In a physical deployment, however, the quality of that enforcement also depends on the hardware capacity, software architecture and operational design.

What is inside a physical firewall appliance?

Physical firewall: functional processing planesConceptual, not tied to a specific appliance chassis. Creator: TechiesJournal. Original conceptual technical figure. Research references: official vendor manuals. Not a vendor architecture reproduction or independent lab result. November 2026. Physical firewall: functional processing planesConceptual, not tied to a specific appliance chassis.Management planeAdmin, policy and updatesControl planeRoutes, HA and decisionsData planePackets and sessionsImplementation detail varies by productCPU, memory, network processor, content engine andinterfaces may share resources or use dedicated hardware.TechiesJournal
Accessible text alternative for this figure

A management plane for admin, policy and updates leads to a control plane for routes, HA and decisions, which leads to a data plane for packets and sessions.

A note states that implementation detail varies by product: CPU, memory, network processor, content engine and interfaces may share resources or use dedicated hardware. The diagram is conceptual.

Figure 1. Logical planes are not necessarily separate chips or independent failure domains.

The common architecture separates a management plane, a control plane and a data plane. These are functional responsibilities rather than a promise that every manufacturer uses three separate chips or processes.

The management plane exposes configuration, administrative access, software updates, certificates, logging settings and operational control. Its security matters because a compromised management interface can alter the traffic policy that the appliance enforces.

The control plane maintains routing and policy information, interface state, HA relationships and other information needed to decide how traffic should be processed. Depending on the product, parts of these responsibilities run in the same software environment as management or packet handling.

The data plane processes packets and sessions. It performs classification, filtering, forwarding, address translation and inspection according to the configured services. Hardware accelerators may take over selected operations once a flow qualifies for a faster path. Other traffic stays on general-purpose processors or goes to dedicated content-inspection components.

Appliance componentTypical responsibilityWhat to verify for a specific model
Physical network interfacesConnectivity, link speed and interface modesPort count, copper/fiber type, supported speed, bypass/aggregation behavior
General-purpose CPU and memoryControl tasks and software inspectionCapacity under enabled security services
Packet processing accelerationEligible forwarding, session or cryptographic workExact processor, supported offloads and exclusions
Inspection engineIPS, application and content analysisWhich protocols, traffic directions and encrypted flows it inspects
Management and storageConfiguration, system images, logs and updatesSeparate access controls, local versus remote logging and upgrade recovery
HA interfaces and state mechanismsHeartbeats, synchronization and takeoverLink requirements, session survival and failover limits

Dedicated hardware does not automatically mean a faster firewall. An appliance with an accelerator that cannot handle the chosen inspection policy may spend significant processing time on a different path. Interface capacity also does not establish inspected traffic capacity.

Management-plane isolation is part of appliance security

Traffic policy is only as dependable as the configuration and administrative authority behind it. A firewall’s management interface should have a deliberately limited exposure path, administrator identities should be controlled, and configuration changes should leave an auditable record. Dedicated management ports can help separate management connectivity, but they do not create security by themselves if the upstream access path is unrestricted.

An engineering design should distinguish the ability to forward user traffic from the ability to administer the device. If the data plane remains healthy but central management is unreachable, some appliances can continue enforcing the last committed policy. If a control-plane service fails, routing updates or session handling may be affected even where hardware remains powered on. The exact failure behavior is implementation-specific. Those distinctions should appear in operating and recovery procedures, rather than being inferred from the presence of redundant power supplies.

The processing path: new flow, inspection and offload

When a session first reaches an appliance, the system may need to classify it, identify the applicable policy, create or locate session state and determine whether additional inspection is required. Once a flow’s treatment is known, some implementations can accelerate subsequent packets by avoiding repeated work on the general-purpose processing path.

Session processing and possible offloadConceptual stages. Offload may retain DPI inspection. Creator: TechiesJournal. Original conceptual technical figure. Research references: official vendor manuals. Not a vendor architecture reproduction or independent lab result. November 2026. Session processing and possible offloadConceptual stages. Offload may retain DPI inspection.Incoming flowSession classificationPolicy and DPIEligibility decisionEligible for accelerationFast packet-processing pathNot eligibleContinued software processingSelected traffic can still reach DPIHardware and policy determine the pathTechiesJournal
Accessible text alternative for this figure

An incoming flow is classified, then a policy and DPI step makes an eligibility decision.

Eligible traffic takes a fast packet-processing path, and traffic that is not eligible continues through software processing. Both can end at a step where selected traffic can still reach DPI, because hardware and policy determine the path. The diagram is conceptual.

Figure 2. Traffic may use a faster packet path while remaining subject to relevant inspection. Eligibility is product-specific.

This diagram is conceptual. It must not be read as the packet pipeline of every NGFW or as proof that deep inspection disappears after offload. Different systems offload different operations, and some accelerated traffic still goes through a security-inspection engine.

Fortinet: distinct network and content acceleration

In its FortiOS 7.6.7 hardware acceleration documentation, Fortinet describes specialized Security Processing Units. The vendor distinguishes network processors that offload high-volume packet processing from content processors that accelerate selected security operations, including content scanning and cryptographic processing. The precise processor combination depends on the appliance model, and some smaller devices do not have all processor classes.

The consequence is architectural, not a universal performance ranking. Two FortiGate models may support a similar feature name yet have different hardware acceleration paths, capacity limits and policies that disqualify traffic from offload. Engineers should inspect both the exact hardware model and the documentation for the installed FortiOS release.

Sophos: SlowPath, FastPath and the inspection engine

Sophos Firewall 23.0 documentation, updated 3 November 2026, describes a SlowPath responsible for initial classification and an eligible FastPath for subsequent processing. The architecture also includes a deep packet inspection (DPI) engine. Depending on the rules, FastPath may bypass repeated kernel work while still delivering packets to the DPI engine when security inspection is required.

This distinction prevents a common misconception. Accelerated forwarding does not always mean content inspection has been abandoned. Conversely, certain policies and traffic types do not qualify for the same acceleration. Sophos documents restrictions involving proxies, particular VPN traffic, QoS/DoS and some interface or protocol arrangements. It also states that active-active HA disables its firewall acceleration and that active-passive deployment has different offload behavior. Those are Sophos-specific implementation details, not universal properties of NGFWs.

On Sophos XGS appliances, model and generation determine whether an Xstream Flow Processor, dedicated cores or software-based FastPath performs eligible work. Thus even products from the same vendor should not be compared using a feature checkbox alone.

Offload can change what a packet capture reveals

Acceleration introduces a diagnostic distinction between packets the device forwards and packets a particular software capture point can observe. In the Sophos Firewall 23.0 architecture documentation, Sophos warns that ordinary capture methods cannot capture FastPath traffic in the same way as SlowPath traffic. It also documents that starting a capture normally redirects traffic to SlowPath temporarily, unless an alternative capture behavior is selected.

That matters for interpreting evidence. The absence of a packet from a software capture does not always prove that it never traversed the appliance. Equally, a capture taken after the device changes processing paths may not represent its normal high-throughput behavior. A reliable verification plan should specify the capture point, whether acceleration is enabled, the device’s counters and logs, and any diagnostic action that modifies packet processing. Those observations are not a substitute for a lab test. They define what an engineer must verify in one.

What does next-generation inspection add?

Basic network filtering distinguishes permitted from denied connections using network criteria and flow state. NGFW services attempt to provide more context about the application and threats traveling across a permitted path. They are complementary to policy enforcement, not a substitute for defining which communication should be allowed.

Application identification

A network connection using TCP port 443 is not proof that the application is a conventional web browser connecting to a web server. Application identification uses signatures, protocol decoding, behavioral clues and sometimes decrypted content to classify traffic with more context than a port number.

Palo Alto Networks’ App-ID documentation describes a sequence in which traffic is evaluated against policy, application signatures and protocol decoders provide additional identification, and decrypted traffic can be classified further when a supported decryption policy applies. The vendor presents this as its own implementation. Its claims about identifying encrypted or evasive applications should not be interpreted as independent evidence that arbitrary encrypted content is visible.

More precise application context can help restrict risky protocols, apply differentiated policies and improve logging. However, application detection depends on signatures, observed traffic and visibility. An application may initially be classified more broadly and then identified more precisely as the session progresses. It is a mistake to assume that every application can be conclusively identified from the first packet.

Intrusion prevention and content inspection

An intrusion prevention system (IPS) inspects eligible traffic against exploit signatures, protocol checks and other detection mechanisms. Antivirus or content scanning may examine file transfers and other supported objects. The appliance needs to reassemble or interpret enough traffic to make the decision, and that work may require different resources from simple packet forwarding.

An inspection engine can produce false positives or miss malicious behavior outside its coverage. Signatures, feature licenses, protocol support and policy selection matter. Products may offer intrusion detection without inline blocking in some modes, or may allow specific flows to bypass inspection. Turning on an IPS profile does not guarantee protection from all exploits.

Deep packet inspection is also not one single function. Network IPS, application identification, file scanning and HTTP-oriented inspection operate on different data and policy criteria. The operational consequence is that enabling one feature does not imply all other analysis is active for the same traffic.

Encrypted traffic changes what the firewall can see

Most sensitive application traffic uses transport encryption. A firewall may observe addresses, ports, connection properties and some protocol metadata without obtaining the decrypted application content. To inspect that content, an organization must use a supported decryption or proxy design and manage its trust implications.

Two frequently discussed approaches are outbound forward-proxy inspection, where the firewall establishes separate client-side and server-side TLS sessions under a trusted enterprise certificate authority, and inbound inspection, where the organization controls the protected service and can provide the credentials or termination arrangement required by the chosen product. These are distinct operational and trust models. Decryption capability must never be inferred solely from an NGFW feature label.

Certificate trust, pinning and mutual TLS

TLS forward proxy: two trust relationshipsIllustrates authorized outbound inspection, not universal TLS visibility. Creator: TechiesJournal. Original conceptual technical figure. Research references: official vendor manuals. Not a vendor architecture reproduction or independent lab result. November 2026. TLS forward proxy: two trust relationshipsIllustrates authorized outbound inspection, not universal TLS visibility.Managed clientTrusts inspection CATLS session ANGFW / TLS proxyDecrypt and inspectRe-encryptRemote serviceServer certificateTLS session BCertificate pinning, mutual TLS or unsupported protocolsmay prevent inspection or require a documented exceptionTechiesJournal
Accessible text alternative for this figure

A managed client that trusts the inspection certificate authority opens TLS session A to an NGFW acting as a TLS proxy, which decrypts, inspects and re-encrypts, then opens TLS session B to the remote service and its server certificate.

Certificate pinning, mutual TLS or unsupported protocols may prevent inspection or require a documented exception. The diagram illustrates authorized outbound inspection, not universal TLS visibility.

Figure 3. Authorized forward proxy decryption introduces a trust boundary and may be incompatible with pinned or mutually authenticated connections.

Forward-proxy decryption requires the client to trust the substitute certificate presented by the inspection system. Applications using certificate pinning may deliberately reject that substitute certificate. Mutual TLS and other authentication designs can also conflict with interception. In its decryption troubleshooting guidance, Palo Alto Networks explains that pinned clients may reject an impersonated certificate. Its decryption exclusion guidance identifies pinning and mutual authentication as reasons some traffic must remain encrypted for technical reasons.

This creates a genuine security trade-off. Blocking an incompatible application may disrupt legitimate business use, but excluding it from decryption reduces content visibility. A policy decision should account for application ownership, privacy expectations, legal requirements and compensating controls. It should not promise transparent interception of every TLS connection.

QUIC and HTTP/3 require a separate inspection decision

Encrypted web traffic is not limited to TCP-based TLS. HTTP/3 commonly operates over QUIC on UDP, so a firewall’s ability to decrypt conventional HTTPS over TCP does not imply the same visibility into QUIC traffic. Palo Alto Networks’ decryption-policy guidance explicitly discusses restricting QUIC when its NGFW decryption policy cannot inspect that traffic and relying on supported TLS paths instead.

The implication is broader than one vendor. Security policy must identify which encrypted transports are permitted, which can be inspected by the deployed product and which remain opaque. Blocking an unsupported transport can change application behavior and availability, so this is an architecture and compatibility choice, not a universal default recommendation. Product capabilities must be checked for the deployed software release.

Decryption also affects throughput and failure behavior

TLS inspection adds handshake processing, certificate operations, encryption/decryption work and further content analysis. Hardware may accelerate some of that processing, but the applicable offload depends on the model, firmware and cryptographic method. The performance of ordinary stateful forwarding is therefore not a substitute for measuring the same device with TLS inspection and IPS enabled.

Inspection can also fail because a certificate is invalid, an application rejects interception, a cipher or protocol is unsupported, or an inspection engine is overloaded. The applicable configured action may be to block, bypass or handle the connection another way. Fail-open versus fail-closed behavior must be verified for the specific inspection feature and failure condition, not asserted as a property of all firewalls.

Why advertised throughput figures are not interchangeable

Firewall datasheets frequently list multiple throughput figures, each measured under different test conditions. The number most prominently advertised is not necessarily the capacity available with the complete security policy enabled.

MetricWhat it measuresWhy it matters
Firewall throughputRate of traffic processed under the manufacturer’s firewall test configurationUsually not equivalent to full NGFW inspection throughput
IPS or threat protection throughputRate with a defined inspection or threat prevention policySensitive to enabled signatures, content and traffic mix
TLS/SSL inspection throughputRate when supported encrypted traffic is decrypted and processedSensitive to cipher suites, handshakes and content inspection
Connections per secondNew sessions that can be established per unit timeCritical for short-lived or bursty workloads
Concurrent sessionsState records supported at the same timeImportant for large user populations and long-lived flows
LatencyAdditional delay through the device under the measured conditionsRelevant to interactive, real-time or latency-sensitive services
Packet forwarding ratePackets processed per secondSmall packets can stress an appliance even at moderate bit rates

As Fortinet’s sizing guidance notes, threat protection throughput can change with SSL/SSH deep inspection. That vendor statement illustrates why a buyer should look for the metric matching the intended policy rather than assume one universal capacity figure.

Comparisons are defensible only when the test methods are compatible. A 1518-byte packet forwarding test, an HTTP transaction workload, a TLS decryption benchmark and a real mixed-traffic corporate environment do not place identical demands on a device. Published figures may also differ in application mix, signature coverage, logging, packet sizes, cipher algorithms, concurrency and whether policy features are enabled.

A workload-based sizing method

Start with the network and security requirements rather than a specific model. Record sustained and peak bidirectional bandwidth, session setup bursts, concurrent connections, average and small-packet traffic, inspection types, encrypted-traffic fraction, VPN requirements, physical interfaces and expected growth.

Next, identify which traffic will undergo each security service. Not every connection necessarily receives decryption, antivirus, IPS or application control. Estimate the load of each relevant traffic class, preserving the assumptions instead of collapsing everything into one headline throughput value.

Finally, request measurements or defensible vendor evidence for the actual combination of policies. Leave capacity headroom for bursts, high-availability failover, signature updates, logging and firmware changes. A numerical safety margin cannot be prescribed from the evidence in this article because it depends on traffic shape and service-level requirements. A validation test against the intended workload is more informative than an unexplained universal percentage.

Availability: the appliance must survive more than a power failure

Physical firewalls often protect critical paths, so their availability design is part of the security architecture. A high-availability pair or cluster can provide a takeover path, but continuity depends on state synchronization, link monitoring, routing behavior, licensing and feature compatibility.

Active-passive and active-active

Firewall HA: takeover is not identical replicationConceptual active-passive layout. Session continuity must be validated. Creator: TechiesJournal. Original conceptual technical figure. Research references: official vendor manuals. Not a vendor architecture reproduction or independent lab result. November 2026. Firewall HA: takeover is not identical replicationConceptual active-passive layout. Session continuity must be validated.Active firewallProcesses protected trafficStandby firewallPrepared for takeoverConfig +supported stateHeartbeat / healthFailover must verify more than device availabilityPaths, session recovery, logs and shared failures need separate checksTechiesJournal
Accessible text alternative for this figure

An active firewall processes protected traffic and sends configuration and supported state to a standby firewall that is prepared for takeover. The standby returns heartbeat and health signals.

A note states that failover must verify more than device availability: paths, session recovery, logs and shared failures need separate checks. The diagram is a conceptual active-passive layout.

Figure 4. An HA pair requires validation of path recovery, supported session continuity, logs and shared failure domains.

In active-passive HA, one device ordinarily processes the protected traffic while another is prepared to take over. In active-active HA, both devices participate in traffic handling, although session ownership and forwarding behavior vary by platform. Active-active does not necessarily mean a simple doubling of inspected throughput.

Palo Alto Networks’ HA mode documentation distinguishes active-passive from active-active designs and describes session handling in each. Its general HA documentation also identifies separate paths for configuration and state synchronization and the events that can trigger failover. Those details must be matched to a particular deployment mode, not generalized across products.

Session synchronization has limits

If the standby unit receives sufficiently current session state, some existing communications can survive failover. Without appropriate state synchronization, clients may have to establish new connections. Even with synchronization, an in-flight handshake, unsupported session type or state update not yet propagated can lead to interruption.

Fortinet documents an explicit performance-versus-continuity trade-off for FortiOS 7.6.6. Its session synchronization guidance describes delaying synchronization for short-lived sessions to reduce overhead, but warns that more sessions may need to restart after failover. This is a concrete example of why a feature called session pickup should not be interpreted as a guarantee that every live exchange is preserved.

Operational failure domains

An HA pair does not automatically eliminate a shared upstream switch, shared power supply, software defect, incorrect configuration pushed to both peers or a routing error. Management access, independent failure domains, heartbeat paths, licensing parity, monitoring and upgrade rollback remain important.

Configuration synchronization is not complete operational replication

A cluster can synchronize certain objects and session state without producing two identical operational records. Palo Alto Networks’ High Availability documentation explains that some data, including log history and certain device-specific management settings, is not shared between peers. It also describes link monitoring, path monitoring and heartbeat conditions that can trigger failover. This means the second appliance’s readiness cannot be determined from configuration sync alone.

An acceptance test should therefore distinguish at least four outcomes: the standby detects the relevant failure, forwarding moves to a surviving path, supported sessions recover or reconnect within an acceptable period, and logs and management remain available for investigation. These checks are a proposed test plan, not test results.

A firmware upgrade can also alter acceleration behavior or require a service restart. For example, Sophos documentation describes operational restrictions when acceleration settings change and a difference in FastPath availability between HA modes. The relevant release documentation should be checked before scheduling maintenance rather than assuming a zero-interruption change.

What an engineering evaluation should actually ask

A useful evaluation separates security capability, effective throughput and operational survivability. A platform that offers every inspection feature but cannot sustain required traffic with those features enabled may not satisfy the deployment. Conversely, a high-throughput appliance configured without the necessary inspection rules may provide less protection than the architecture assumes.

Before committing to a device, engineers should be able to answer these questions:

  1. What traffic must cross the appliance, and what controls are required for each traffic class?
  2. Which model and installed software version support those controls?
  3. Which operations use the general CPU, hardware acceleration or DPI engine?
  4. Which flows cannot be offloaded, and what happens when an accelerated path is unavailable?
  5. Which encrypted applications can be inspected, and which need documented exclusions?
  6. What throughput, session-rate and latency evidence matches the intended policy mix?
  7. How do failover, session synchronization and upgrade behavior affect critical applications?
  8. Can the organization operate, monitor and recover the appliance safely throughout its lifecycle?

The right product is the one that meets those requirements with a defensible operating model. An NGFW label, a processor name or a single throughput number is insufficient on its own.

References and further reading

Report a correction

Corrections go to the editor and are never published automatically. No account needed.