data-page="blog">
← All articles
Networking

How Do You Like Them SD-WAN Appliances?

How Do You Like Them SD-WAN Appliances?

Almost every networking appliance with two WAN ports is now described as an SD-WAN appliance.

Some provide basic internet failover. Others distribute sessions across multiple connections. A smaller number can continuously assess link quality, identify applications, steer traffic, mitigate packet loss, bond multiple WAN links and apply security policies across every branch.

These are very different capabilities, even though they may all carry the same SD-WAN label.

For an IT manager, IT director or CTO evaluating SD-WAN appliances, the important question is not whether a product supports SD-WAN. The better question is: What exactly does the appliance do with multiple WAN connections?

The answer also reveals something important about the OEM. Basic failover and load balancing can often be assembled using standard routing functions. Application awareness and packet-level bonding require much deeper ownership of the product, including its traffic-classification engine, data plane, network protocols, cloud infrastructure and central orchestration platform.

The 5 tiers of SD-WAN

TIER 5 Comprehensive SD-WAN Packet-level bonding, single-flow aggregation, loss mitigation, integrated security Edge Ultra, Edge Extreme TIER 4 Centrally Orchestrated SD-WAN Zero-touch provisioning, central policy, network-wide monitoring Edge Pro TIER 3 Application-Aware SD-WAN Application identification, link-quality monitoring, dynamic path selection TIER 2 Policy-Based WAN Rules by destination or traffic type, prioritisation, latency awareness Edge Lite Series TIER 1 Basic Multi-WAN Multiple links, keepalive monitoring, failover, basic load balancing Each tier includes everything in the tiers below it.
Every tier is sold as "SD-WAN". What differs is what the network can actually do. Capabilities and model mapping are from the NetSense SD-WAN tiers guide below.
PDF
Free download
The 5 Tiers of SD-WAN: a visual guide
What each tier of SD-WAN can do, from basic failover to packet-level bonding, and which NetSense Edge model fits each tier. PDF, 1.1 MB.
Download the guide

Multi-WAN failover

Multi-WAN failover is the simplest form of WAN resilience.

The appliance uses one connection as the primary WAN and keeps another connection on standby. If the primary connection becomes unavailable, traffic is moved to the backup.

For example, a branch may use:

Failover protects the branch from a complete link outage. However, basic implementations may only check whether the gateway is reachable. A connection can remain technically online while suffering from high latency, packet loss, DNS failure or an upstream internet problem.

A better implementation checks multiple destinations and evaluates whether the path is actually usable.

Failover provides continuity, but it does not use the combined bandwidth of all available connections.

Weighted load balancing

Weighted load balancing distributes traffic across multiple WAN links according to configured proportions.

Suppose a branch has a 200 Mbps fibre connection and a 100 Mbps broadband connection. The administrator may assign twice as much traffic to the fibre connection as to the broadband connection.

This helps use both links and prevents one connection from carrying the entire branch load.

The weights are normally applied when new sessions are created. They do not usually divide the packets of an individual session across both connections.

Weighted load balancing can increase the total bandwidth available to the organization, but it does not necessarily increase the speed available to one user, download or TCP connection.

Per-session load balancing

Per-session load balancing assigns each new network session to a WAN link.

One employee’s video call may use WAN 1, while another employee’s software download uses WAN 2. A third session may use WAN 1 again, depending on the balancing algorithm and configured weights.

Once a session has been assigned to a path, it usually remains on that path for its lifetime. This is necessary because changing the source IP address in the middle of a TCP session can cause the remote server, firewall or NAT device to reject the traffic.

Per-session balancing can make effective use of multiple connections when many users and applications are active. It is less useful when the organization expects a single large session to consume the combined capacity of all WAN links.

Policy-based routing

Policy-based routing selects a WAN path according to rules defined by the administrator.

Policies can consider attributes such as:

An organization might route payment traffic through an MPLS circuit, general web browsing through broadband and software updates through a lower-cost internet connection.

Policy-based routing provides useful control over traffic placement. However, it is generally dependent on rules created by the network administrator. It does not necessarily understand the application behind the traffic, nor does it automatically choose a better path when latency, jitter or packet loss changes.

Application-aware routing

Application-aware routing makes path decisions based on the actual application, not merely an IP address or port number.

This is more difficult than it sounds.

Modern applications use HTTPS, cloud hosting platforms, shared content-delivery networks and frequently changing IP addresses. Several unrelated applications may use the same destination network or TCP port. A simple port-based rule cannot reliably distinguish a business application from general web traffic merely because both use port 443.

A credible application-aware SD-WAN platform needs deep and continuously maintained knowledge of:

This intelligence must be supported by a classification engine that can combine domain, IP, protocol, session and traffic information. It also needs regular updates because cloud services continually change their infrastructure.

The quality of application-aware routing therefore depends heavily on the OEM’s technical ownership and operational knowledge.

An OEM that primarily rebrands a generic appliance may be able to provide port-based rules or a fixed list of domains. It may still describe this as application-aware routing. But if the OEM does not control the classification engine, maintain the application intelligence and integrate it with the routing data plane, it will struggle to identify applications accurately or adapt when those applications change.

Real application awareness is not a checkbox. It is an ongoing product and engineering function.

Dynamic path selection

Dynamic path selection continuously measures the condition of each WAN link and selects a path based on real-time performance.

The appliance may monitor:

A voice or video application may require low latency and jitter, while a backup process may tolerate delay but need higher throughput.

Instead of asking whether a WAN link is simply up or down, dynamic path selection asks whether the link is good enough for a particular application.

This works best when application awareness and path measurement are part of the same platform. The appliance must first identify the traffic correctly and then understand what network conditions that application can tolerate.

If the preferred path deteriorates, new sessions can be directed to another connection. More advanced architectures may also move active traffic without interrupting the application. Doing this reliably requires an overlay that can preserve addressing, encryption and session state.

Packet-level WAN bonding

Packet-level WAN bonding distributes packets belonging to the same traffic stream across multiple WAN connections.

This is fundamentally different from load balancing.

With load balancing, complete sessions are generally assigned to different WAN links. With packet-level bonding, packets from one session can travel through multiple links and be reassembled at a remote endpoint.

That remote endpoint may be a data-centre gateway, cloud gateway or SD-WAN point of presence. It receives packets from the different paths, restores their correct order and forwards the reconstructed traffic towards its destination.

A functional WAN-bonding architecture must handle:

Packet-level bonding can improve link utilization and may increase the bandwidth available to a single application. However, it requires control at both ends of the bonded path. It cannot normally be delivered by installing a multi-WAN appliance at the branch without a compatible endpoint elsewhere.

This is also where many OEMs reach a technical ceiling.

Failover and per-session load balancing can be implemented using standard routing, NAT and connection-tracking functions. Packet-level bonding requires a purpose-built data path, a tunnelling protocol, packet scheduling, sequencing, encryption and remote reassembly infrastructure.

The OEM also needs to operate or integrate the remote gateways through which bonded traffic will pass. It must be able to diagnose behaviour across the branch appliance, multiple internet connections and the reassembly endpoint.

If the vendor does not own the appliance software, bonding engine, protocol and remote infrastructure, its ability to improve performance or fix difficult field problems will be limited. A reseller or brand owner cannot easily add this capability when the underlying product supplier has not designed for it.

Parity-based packet-loss mitigation

Parity-based packet-loss mitigation sends calculated recovery data alongside the original packets.

The sender groups a number of packets and creates one or more parity packets from them. If a limited number of packets are lost in transit, the receiving endpoint may reconstruct the missing information without waiting for the original sender to retransmit it.

This can be valuable for:

Parity does not eliminate every instance of packet loss. Recovery depends on how much additional parity data is transmitted and how many packets are lost within a group.

More parity can improve recovery, but it also consumes additional bandwidth. The SD-WAN system must decide how much additional data to send, which paths should carry it and whether the application will benefit from it.

Parity-based mitigation is particularly useful when combined with packet-level bonding. Original and recovery packets can be distributed across different WAN connections so that a short disruption on one path does not immediately affect the application.

As with bonding, this requires engineering at both endpoints. A basic multi-WAN appliance cannot reconstruct lost packets unless the sender and receiver use a coordinated packet format and recovery mechanism.

Central orchestration

Managing a few branch appliances individually may be possible. Managing hundreds of them separately is not.

Central orchestration provides one place to define and distribute configurations across the SD-WAN environment. It can manage:

Central orchestration is one of the capabilities that separates an operational SD-WAN platform from a collection of independent multi-WAN routers.

It also exposes the level of product ownership behind the solution. A deeply integrated orchestrator can understand application identities, overlay topology, path quality, bonding performance and security policy as parts of the same system.

A lightly integrated management portal may only push configuration files and display basic device status. That can still be useful, but it is not the same as controlling an adaptive SD-WAN fabric.

Zero-touch deployment

Zero-touch deployment allows a branch appliance to be installed without sending a network engineer to configure it locally.

The appliance can be shipped to the branch, connected to power and attached to an internet connection. It then contacts the central orchestration platform, verifies its identity and downloads the appropriate configuration.

A good zero-touch deployment process should securely handle:

Zero-touch deployment becomes especially important for banks, NBFCs, retailers, healthcare networks and other organizations with many distributed locations.

It reduces installation time, limits configuration errors and makes large-scale rollouts more practical. It should, however, be evaluated as a security-sensitive provisioning system, not simply as automatic configuration download.

SD-WAN integrated with security and SASE

Branch connectivity and security are closely connected. Traffic leaving a branch may need firewall inspection, application control, intrusion prevention, web filtering, malware protection and access policies.

An SD-WAN appliance integrated with security can apply these controls while also selecting the appropriate path.

A SASE architecture extends this model by delivering networking and security capabilities through distributed cloud points of presence. Users and branches connect to the appropriate point of presence, where policies can be applied before traffic reaches the internet, a SaaS platform or a private application.

An integrated SD-WAN and SASE architecture may include:

Integration should be assessed carefully. Running SD-WAN and security features on the same appliance does not automatically mean they share the same application intelligence, policy framework, management system or traffic context.

The real value appears when the platform can identify an application once and use that understanding for routing, bandwidth management, security inspection and reporting.

That level of integration is difficult when the SD-WAN appliance, firewall, application database, cloud gateway and orchestration platform come from separate suppliers over which the OEM has little technical control.

Do two 100 Mbps connections create a 200 Mbps connection?

Not automatically.

This is one of the most commonly misunderstood claims in multi-WAN and SD-WAN discussions.

Consider a branch with two 100 Mbps internet connections. A conventional load-balancing appliance can use both connections and provide up to 200 Mbps of aggregate capacity across many simultaneous sessions.

For example:

The organization can consume traffic through both connections at the same time. But a single download or application session will usually remain limited by the capacity of the WAN link to which it has been assigned.

This happens because a TCP connection is identified by its source and destination addresses and ports. When a session passes through a WAN connection, network address translation normally gives it the public IP address associated with that connection.

If packets suddenly leave through the second WAN, they may appear to come from a different public IP address. The destination server will not treat them as part of the original TCP connection.

Changing paths can also cause packets to arrive significantly out of order. TCP may interpret this as congestion or loss, reduce its transmission rate and negate the expected performance benefit.

For these reasons, normal multi-WAN load balancing keeps each TCP flow pinned to one WAN.

How packet-level bonding changes this

To give one session access to the combined capacity of multiple links, the SD-WAN system needs packet-level bonding and reassembly.

The branch appliance first places the original traffic inside an encrypted overlay. It can then distribute the overlay packets across both WAN connections.

A remote bonding endpoint receives packets from the two paths, restores their correct order, handles missing or delayed packets and reconstructs the original traffic. The destination application sees one stable connection from the remote endpoint rather than two unrelated internet connections from the branch.

In ideal conditions, two 100 Mbps links may provide throughput approaching their combined capacity. In practice, usable performance will be lower because of:

If one connection has much higher latency than the other, the bonding engine must decide how quickly to send traffic through each path and how long to wait for delayed packets. Poor scheduling can delay the entire stream and reduce performance.

Meaningful WAN bonding therefore requires more than sending alternate packets through alternate connections. It requires continuous path measurement, intelligent scheduling, sequence management and a remote endpoint designed to reassemble the traffic.

This capability is difficult to add through branding, configuration or a management interface. It has to be built into the architecture of the product.

Two 100 Mbps links: what one session actually gets

0 100 Mbps 200 Mbps One session, load balancing Pinned to one WAN link Up to 100 Mbps Many sessions, load balancing Spread across both links Up to 200 Mbps in total One session, packet-level bonding Overlay, reassembled at a bonding endpoint Approaching 200 Mbps, less overhead Branch with two 100 Mbps internet links. Bonded figure is illustrative.
Before accepting "200 Mbps", ask: for the whole branch, or for one application session?

Why some OEMs stop at failover and load balancing

The different SD-WAN capabilities can be viewed as progressively deeper levels of engineering.

Failover requires link monitoring and route changes. Weighted and per-session load balancing add traffic distribution and connection tracking. Policy-based routing adds administrator-defined decisions.

Application awareness requires maintained intelligence about domains, IP addresses, protocols and traffic behaviour. It must be connected to a classification engine that can accurately identify applications as they evolve.

Packet-level bonding goes deeper still. It requires ownership of the packet-processing path, overlay protocol, scheduling logic, encryption, parity mechanism, reassembly engine and remote gateway infrastructure.

This explains why some OEMs cannot move beyond failover, load balancing or basic policy-based routing. The limitation may not be the appliance hardware. It may be the OEM’s lack of control over the underlying software and architecture.

If the product is assembled from generic components or sourced as a finished white-label appliance, the OEM may be able to change branding, user-interface elements and configuration options. It may not be able to alter how packets are classified, scheduled, encrypted or reassembled.

This matters after deployment as well. When an application is misclassified or bonded traffic performs poorly across unequal links, resolving the problem requires engineers who understand and control the relevant parts of the system.

Deep product ownership does not guarantee that every feature will work well. But without that ownership, developing, maintaining and supporting advanced SD-WAN capabilities becomes considerably more difficult.

What should an IT leader ask an SD-WAN vendor?

When evaluating an SD-WAN appliance, ask the vendor to describe precisely how traffic is handled.

Useful questions include:

  1. Does the appliance provide only failover, or can it use multiple active WAN connections?
  2. Is load balancing performed per session or per packet?
  3. Can an individual TCP session use more than one WAN link?
  4. If packet-level bonding is supported, where does reassembly take place?
  5. Does the OEM own and develop the bonding protocol and reassembly engine?
  6. What happens when bonded links have different latency and packet-loss characteristics?
  7. How does the platform identify applications that share HTTPS and cloud infrastructure?
  8. Who maintains the domain, IP and traffic intelligence used for application classification?
  9. How frequently is that intelligence updated?
  10. Can routing decisions consider application identity, latency, jitter and packet loss together?
  11. Can existing sessions survive a path failure?
  12. Are SD-WAN and security policies managed from the same platform?
  13. Does zero-touch deployment include secure device authentication?
  14. What remote infrastructure is required for bonding and cloud-delivered security?
  15. Can the vendor demonstrate performance with one TCP session as well as multiple sessions?

The distinction between aggregate bandwidth and single-session bandwidth should be visible in the test methodology.

A multi-session speed test may show traffic using both WAN links, but it does not prove that a single application flow is bonded across them. Similarly, displaying an application name in a dashboard does not prove that application identity is accurate enough to drive routing and security decisions.

The right SD-WAN depends on the outcome you need

There is nothing inherently wrong with a multi-WAN appliance that provides failover and load balancing. For many branches, those capabilities may be sufficient.

The problem begins when different technologies are presented as if they deliver the same result.

Failover improves availability. Weighted and per-session load balancing improve aggregate link utilization. Policy-based routing provides administrative control. Application-aware routing makes decisions based on accurately classified traffic. Dynamic path selection responds to changing link quality. Packet-level bonding can combine paths for an individual flow. Parity-based mitigation can help applications operate across lossy connections. Central orchestration and zero-touch deployment make the network manageable at scale. Integrated security and SASE extend consistent protection across branches, cloud applications and remote users.

The more advanced capabilities require progressively deeper ownership of the product. They cannot be created merely by adding the SD-WAN name to a multi-WAN router.

An IT leader should therefore evaluate both the appliance and the engineering organization behind it. Ask what the system does, how it does it, which components the OEM owns and who can fix the system when real-world networks behave differently from a laboratory demonstration.

Before accepting that two 100 Mbps connections will provide 200 Mbps, ask one final question:

Two hundred megabits for the entire branch, or for one application session?

The answer will tell you a great deal about both the SD-WAN appliance and the OEM offering it.

PDF
Free download
The 5 Tiers of SD-WAN: a visual guide
What each tier of SD-WAN can do, from basic failover to packet-level bonding, and which NetSense Edge model fits each tier. PDF, 1.1 MB.
Download the guide

See it in your environment

Talk to a NetSense security architect about applying this to your stack.

Get a Demo