This runbook helps network and system administrators add a second WAN without choosing paths blindly. It explains Performance SLA, traffic rules, session behavior and failover, then shows how to check the result on the target FortiGate model and release.
The configuration example uses FortiOS 7.6.3, with core syntax also available in the 7.4.6 CLI reference. Check the options supported by your device model and patch, then run acceptance tests for that combination.
Introduction
Before adding the second WAN, decide what counts as a usable internet path. A lit port establishes the local link, and a responding gateway does not establish the health of the whole route. The ISP gateway may answer ping while its upstream drops packets or delays voice calls. Availability improves when FortiGate can distinguish these conditions.
Two static routes with different distances usually create primary and backup paths; the lower distance wins. This alone provides neither load balancing nor quality detection. Simple ECMP supplies equal-cost routes but no independent quality policy for Microsoft 365, voice or large downloads. Traditional routing can use a link monitor; the limitation discussed here is ECMP or static routing without monitoring and service policy.
This guide uses the following operational architecture:
FortiGate → SD-WAN → Multiple ISP → Performance SLA → Traffic Steering → Source-Destination Load Balancing → Automatic Failover
SD-WAN ties path selection to traffic type and link quality. Firewall policy still authorizes traffic, and NAT and routing must be correct; SD-WAN does not replace them. Official SD-WAN Rules documentation
What is load balancing in FortiGate?
Here, load balancing distributes connections or groups of connections: one group of users might leave through WAN1 and another through WAN2. FortiGate tracks each session using information such as addresses, ports and protocol. The chosen route and NAT translation belong to that session's state, which matters when interpreting a path change.
This is not ISP bonding. A normal TCP session is not split across WANs and cannot reach the sum of 200 and 100 Mbps. Many independent connections can use both links. Actual speed depends on the selected WAN, destination server, RTT, congestion, FortiGate processing capacity and security inspection.
A download manager opening several independent TCP connections may use both WANs with some algorithms. With source-destination IP hashing, connections sharing an IP pair usually take one WAN. Even a multi-connection download therefore need not use both links simultaneously.
Load balancing versus failover
| Concept | Purpose | Example behavior |
|---|---|---|
| Load Balancing | Use multiple paths to distribute load | New connections are spread across WAN1 and WAN2 |
| Failover | Continue service after path failure or degradation | New connections use an alternative link |
There are two separate decisions: which link receives a new connection, and which path is no longer suitable for service. Without a sound health check, balancing can send new connections to a broken internet path. Conversely, failover can operate while only one link carries normal traffic. The proposed design uses both decisions together.
Active/active and active/passive
These terms refer to Internet links here and must not be confused with FortiGate HA cluster modes.
Active/active: both WANs carry some traffic. This suits organizations needing both ISPs' capacity whose applications tolerate different egress IPs across independent connections. Sensitive traffic can still have dedicated rules.
Active/passive: WAN1 is primary and WAN2 is backup. This is more predictable when WAN2 is metered, has limited capacity or business applications need a stable egress IP. For quality-based failover, use an SLA strategy with load balancing disabled and appropriate priority/cost; different route distances alone do not detect brownouts.
In either model, the remaining link must carry essential traffic. If normal use is 180 Mbps, a 100 Mbps WAN2 cannot carry everything at the same quality during an outage. QoS and backup throttling during failures belong in the design.
Load-balancing methods
FortiOS distinguishes the implicit rule's algorithm from the strategy/hash in an explicit SD-WAN service rule. Their options differ. Session-based and weight-based are not separate implicit algorithms: the GUI Sessions option uses weight-based. Selecting the implicit SD-WAN algorithm
Source IP Based
A hash is built from the source IP. Connections from one client to different destinations map to the same path while the eligible links and selection state remain unchanged.
- Advantage: more stable egress IP per client and simpler behavior for applications sensitive to IP changes.
- Limitation: a heavy client or proxy representing many users behind one IP can saturate one WAN.
- Use: many clients with roughly similar loads; services needing per-user egress stability.
- Persistence: at source-IP level; no permanent guarantee after failover or topology changes.
- Enterprise: usable, but evaluate proxies, internal NAT and small source populations with uneven loads.
Source-Destination IP Based
The hash uses the source/destination IP pair. A client can map different destinations to different WANs while its connections to a particular IP share one path.
- Advantage: a useful balance between load distribution and source/destination stability.
- Limitation: hashing alone does not understand actual capacity, session size or application sensitivity; a few large flows can skew distribution.
- Use: general Internet alongside SLA and application-specific rules.
- Persistence: at IP-pair level; mappings may change when eligible links change.
- Enterprise: suitable for general traffic with explicit exceptions for sensitive applications.
Session Based
Distribution decisions apply to new sessions. The implicit rule uses session count and member weights; explicit rules also offer round-robin distribution of new connections. These behaviors are not identical.
- Advantage: even many connections from one source to one destination can be distributed across WANs.
- Limitation: one application's sessions may use different public IPs; a short session and a heavy download count equally as connections.
- Use: few clients with many independent connections and applications tolerating multiple egress IPs.
- Persistence: a session retains its path, but the application's next connection may take another.
- Enterprise: suitable for selected traffic; test banking, SSO and IP-allowlisted services.
Weight Based
Implicit member weights set approximate shares of session count. Weights 2 and 1 imply statistical shares near two-thirds and one-third; they do not guarantee bandwidth shares.
- Advantage: roughly aligns connection distribution with unequal links.
- Limitation: session sizes vary; connection-count ratios need not match Mbps ratios.
- Use: unequal capacities with many independent sessions.
- Persistence: like session-based; source/destination pair stability is not a general requirement.
- Enterprise: valid when distinguished from explicit-rule hashing and measured against real loads.
Volume Based
Distribution targets measured traffic volume. The implicit rule uses measured-volume-based and volume-ratio.
- Advantage: more sensitive to transferred data than session counting.
- Limitation: past volume does not guarantee current capacity or application quality; an already selected long flow is not split across WANs.
- Use: bulk traffic with unequal capacities, after testing the device model and offload behavior.
- Persistence: sessions usually continue on their selected path; later sessions may choose differently.
- Enterprise: usable with monitoring and real tests, not as an SLA replacement.
Spillover
The first link is used until a utilization threshold; additional sessions use the next. Its implicit CLI option is usage-based.
- Advantage: prefers cheaper Internet and uses the second WAN as load increases.
- Limitation: requires correct thresholds and hardware/offload considerations; it does not create equal distribution.
- Use: a fixed-cost primary ISP and a metered secondary ISP.
- Persistence: existing sessions generally do not move merely because the threshold is exceeded.
- Enterprise: useful for cost-driven scenarios, not a universal recommendation.
Fortinet's 7.4.1 spillover example requires disabling auto-asic-offload in the policy; measure the processing cost first. Implicit rule
Best Quality / SLA Based
Best Quality ranks paths by a chosen metric such as latency or jitter. Lowest Cost (SLA) first prefers SLA-compliant links, normally using cost and preference for selection. Enabling load balancing in SLA mode distributes traffic among compliant links; cost is not a distribution weight in this mode.
- Advantage: responds to brownouts and real path quality even while the port stays up.
- Limitation: results depend on targets, probe protocol, thresholds and fallback policy.
- Use: VoIP, business applications, VPN and general Internet with suitable SLAs.
- Persistence: selects a new connection's path; it does not guarantee seamless migration of existing connections.
- Enterprise: the foundation of this design; select appropriate metrics for each service class.
Best quality strategy; Load balancing strategy
Algorithm comparison
Figure 4 — Path-selection criteria versus persistence scope; path quality is a separate layer from the distribution algorithm.
| Method | Criterion | Main advantage | Main limitation | Persistence across connections | Enterprise suitability |
|---|---|---|---|---|---|
| Source IP | Source hash | Stable egress per client | Heavy-source load concentration | High at source level | Suitable with load assessment |
| Source-Destination | IP-pair hash | Balance of stability and distribution | Unaware of flow size | High for the same pair | Suitable for general Internet |
| Session | New connection/session count | Distributes multiple connections | Different egress IPs for one application | No IP-pair guarantee | Suitable for selected traffic |
| Weight | Session-count ratio | Supports unequal capacities | Mbps is not guaranteed | Like session-based | Suitable with monitoring |
| Volume | Data-volume share | More sensitive to usage | Measurement-dependent response | Retained within a session | Requires testing |
| Spillover | Utilization threshold | Controls secondary-WAN cost | Sensitive to threshold/offload | Retained within a session | Suitable for cost-driven use |
| Best Quality | Quality ranking | Improves sensitive-service paths | Depends on probes | New selections depend on quality | Suitable for critical applications |
| SLA + Load Balance | SLA then hash | Quality and distribution together | Requires clear fallback | Depends on selected hash | Recommended base architecture |
Why use source-destination hashing for general traffic?
Suppose the client is 10.10.10.25 and destination 8.8.8.8. FortiGate hashes this pair and maps it to an eligible member. We cannot assert WAN1 without observing the device; assume WAN1 is selected here.
10.10.10.25 → 8.8.8.8 → Hash pair A → WAN1
10.10.10.25 → 1.1.1.1 → Hash pair B → WAN1 or WAN2
10.10.10.26 → 8.8.8.8 → Hash pair C → WAN1 or WAN2
Connections for 10.10.10.25 / 8.8.8.8 map to the same WAN with a stable healthy-member set and the same rule. Changing the source port alone does not drive this hash. With session-based distribution, a new connection for the same pair may take another WAN.
This helps networks with varied users and destinations but does not guarantee website session persistence. A website may use multiple IPs, CDNs and separate authentication services; clients may also change IP. Applications needing consistent egress should have a dedicated preferred-path rule with defined fallback.
Why SD-WAN?
Define each traffic group's needs before choosing its path. Conventional routing finds an available route; SD-WAN can also evaluate whether that route has suitable quality for the traffic. This lets you distribute general internet access, send voice through a low-jitter path and use a cheaper link for backups.
SD-WAN is a WAN-selection policy engine; application identification, SLAs and routes must align. Application Control identification may complete after a connection begins; a port match alone does not prove application identification. Dynamic application steering
Prerequisites
Use a supported release, a tested backup, administrative access, a maintenance window when required, and a rollback plan. Validate DNS, time synchronization, firewall rules, storage, and monitoring before implementation.
Proposed architecture
| Component | Interface | Specifications | Example FortiGate IP | Gateway |
|---|---|---|---|---|
| WAN1 | wan1 |
ISP-1; 200 Mbps | 203.0.113.2/30 |
203.0.113.1 |
| WAN2 | wan2 |
ISP-2; 100 Mbps | 198.51.100.2/30 |
198.51.100.1 |
| LAN | internal |
10.10.10.0/24 |
10.10.10.1/24 |
User gateway: 10.10.10.1 |
| SD-WAN Zone | internet-zone |
Contains WAN1 and WAN2 | None | Each member's gateway |
WAN addresses use TEST-NET documentation ranges and are not routable on the real Internet. Replace them with actual ISP-assigned IP/mask/gateway values. 1.1.1.1 and 8.8.8.8 are the example's only real public addresses, intentionally used as probe targets.
Internet
/ \
ISP-1 ISP-2
200 Mbps 100 Mbps
| |
wan1 wan2
\ /
FortiGate
[SD-WAN engine]
|
internal
|
10.10.10.0/24
SD-WAN is not a physical link between FortiGate and the LAN; its diagram label represents the logical path-selection engine inside FortiGate.
Figure 2 — Two independent WANs belong to internet-zone; LAN users use FortiGate's internal gateway.
Execution prerequisites
The CLI scenario assumes IPv4, NAT/route mode, one VDOM named root, profile-based NGFW and Central NAT disabled. Interfaces wan1, wan2 and internal must actually exist. Some models use different LAN names or switch-member WANs; inspect the real structure before pasting.
Back up configuration, retain console or independent management access, and inspect routes, policy routes, VIPs, DHCP/PPPoE and policies referencing the WANs. Adding an interface with incompatible references may fail. Member IDs 1 and 2 and rule ID 10 must be unused; editing an existing ID changes that object.
This example uses two static WANs. Simply retaining fixed gateways when converting to DHCP/PPPoE is incorrect; design their automatic gateways/default routes separately. Client DNS must already be defined; this guide does not redefine organizational DHCP/DNS servers.
Performance SLA
Treat the health check as a path sensor: it measures each WAN's quality to an external destination. The performance SLA then compares those measurements with the service requirements. Together, these steps establish whether a path is suitable for that service.
| Metric | Operational definition | Example value |
|---|---|---|
| Latency | Probe round-trip delay in this ping design | 150 ms |
| Jitter | Variation in sample delay | 30 ms |
| Packet Loss | Percentage of unanswered probes in the measurement window | 5% |
These values are examples. Five percent loss can be unacceptable for voice; 150 ms may be unattainable on an intercontinental path even normally. Record baselines by location, peak period, service and ISP, then derive thresholds from actual SLOs.
Target selection and avoiding false positives
General health checks use 1.1.1.1 and 8.8.8.8 on independent networks. Pinging public DNS does not prove HTTPS or Microsoft 365 works. Anycast, rate limiting and destination route changes also affect results.
With two health-check servers, FortiOS probes the first and uses the second if the first becomes unavailable; this is not simultaneous voting. A reachable but slow first server can fail SLA, and a second server alone does not resolve that. In production, add independent service-specific checks such as HTTPS to a controlled endpoint, and define each rule's selection/fallback logic. Link health monitor
Do not treat multiple targets as AND, OR or majority without checking SLA combination behavior. ISP health and application health differ; a global SaaS outage should not necessarily flap all organizational Internet traffic.
Alive/dead differs from SLA pass/fail
Dead: probes receive no response and health-check failure handling activates. Alive but SLA failed: the path responds below required quality. A default route may still exist while the rule prefers another path.
If WAN1 fails SLA and WAN2 passes, new general-rule connections should use WAN2. If both are alive but outside SLA, this strategy can use both to preserve connectivity; SLA is not an absolute traffic prohibition. Design and test fail-closed requirements separately. Lowest cost SLA — all links outside SLA
Figure 3 — Assess link health and quality, then use the service rule and distribution algorithm to select a new session's path.
GUI implementation
The documented path in recent FortiOS is:
Network → SD-WAN → SD-WAN Zones
Network → SD-WAN → Performance SLAs
Network → SD-WAN → SD-WAN Rules
In the checked 7.4/7.6 documentation, Performance SLAs is a tab inside Network → SD-WAN, rather than a standalone Network → Performance SLA menu. Control placement varies by patch and feature visibility. Official GUI SLA guide
- Under Network → Interfaces, configure WAN and LAN addresses according to the table.
- Under Network → SD-WAN, create internet-zone and add wan1 and wan2 with their gateways.
- On Performance SLAs, create Internet-Probe with Ping, the two example servers and both WAN participants. Enable SLA targets with the table's thresholds.
- On SD-WAN Rules, create General-Internet with the LAN source, destination all and Lowest Cost (SLA). Select the SLA and both members, and enable load balancing.
- Under Network → Static Routes, create a default route through internet-zone.
- Under Policy & Objects → Firewall Policy, allow internal to internet-zone from the LAN, using outgoing-interface NAT and logging.
- Complete the hash through CLI. Documentation describes this strategy's GUI balancing option as round-robin. Enabling GUI load balancing does not imply source-destination hashing; set hash-mode source-dest-ip-based in the explicit rule in the next step.
In FortiOS 7.6, the Performance SLAs tab inside SD-WAN displays latency, jitter and packet loss for both WANs. Compare your device with an actual capture of that release; the diagrams in this guide explain the design and are not product screenshots.
CLI implementation
Version compatibility and execution order
From FortiOS 7.4.1 onward, this scenario uses set mode sla with set load-balance enable, as shown in the 7.6.3 and 7.4.6 references. Convert older set mode load-balance examples before using them on a newer release, and inspect command help for the installed patch. Strategy changes
Run blocks 1–6 in order in the target VDOM after resolving incompatible WAN references. Together they configure the defined scenario; the weighted and steering sections are alternatives or extensions.
1. Interface addresses
config system interface
edit "wan1"
set mode static
set ip 203.0.113.2 255.255.255.252
set role wan
set status up
next
edit "wan2"
set mode static
set ip 198.51.100.2 255.255.255.252
set role wan
set status up
next
edit "internal"
set mode static
set ip 10.10.10.1 255.255.255.0
set role lan
set status up
next
end
config system interface opens the interface table; edit selects an existing interface. mode static chooses static addressing, ip sets address/mask, role sets its display/management role, and status up enables it administratively without proving physical connectivity or Internet health. next saves the object; end exits. This block intentionally leaves management access unchanged; access the device only from an authorized management network. Official Interface CLI Reference
2. LAN address object
config firewall address
edit "LAN-10.10.10.0_24"
set subnet 10.10.10.0 255.255.255.0
next
end
config firewall address opens the address table; edit names the object and subnet sets the matched network. WAN selection and firewall policies use this object. Official Address Object example
3. SD-WAN zone and members
config system sdwan
set status enable
set load-balance-mode source-dest-ip-based
config zone
edit "internet-zone"
next
end
config members
edit 1
set interface "wan1"
set zone "internet-zone"
set gateway 203.0.113.1
next
edit 2
set interface "wan2"
set zone "internet-zone"
set gateway 198.51.100.1
next
end
end
status enable activates SD-WAN. load-balance-mode selects the implicit algorithm, not an explicit rule's hash-mode. config zone creates the logical zone. Under config members, edit is the member ID, interface the real interface, zone its membership, and gateway the ISP next hop. Set gateways per member rather than inventing a shared gateway for two ISPs. CLI Reference 7.4.6
4. Health check and SLA target
config system sdwan
config health-check
edit "Internet-Probe"
set server "1.1.1.1" "8.8.8.8"
set protocol ping
set members 1 2
set interval 1000
set failtime 5
set recoverytime 10
set update-static-route enable
config sla
edit 1
set latency-threshold 150
set jitter-threshold 30
set packetloss-threshold 5
next
end
next
end
end
server defines two primary/secondary targets. protocol ping uses ICMP, and members 1 2 monitors both WANs. interval 1000 sets probe spacing in milliseconds. failtime 5 and recoverytime 10 require consecutive failures/successes for health-state changes; they are not latency/jitter thresholds.
update-static-route enable allows the check to affect associated static routes when down; exceeding a latency SLA is not the same as route removal. Under config sla, target 1 sets latency/jitter in ms and packet loss as a percentage. The correct CLI keyword is packetloss-threshold.
Failover is not exactly five seconds: timeouts, fallback servers, measurement windows and actual link state affect timing. Ten successful recovery probes do not guarantee ten seconds of stable quality for all applications. CLI Reference 7.6.0
5. SD-WAN service rule with source-destination hash
config system sdwan
config service
edit 10
set name "General-Internet"
set mode sla
set src "LAN-10.10.10.0_24"
set dst "all"
set load-balance enable
set hash-mode source-dest-ip-based
config sla
edit "Internet-Probe"
set id 1
next
end
set priority-members 1 2
next
end
end
config service opens SD-WAN rules; edit 10 sets the ID and name sets its name. mode sla enables SLA-based selection. src matches the LAN and dst all its destinations. load-balance enable distributes among eligible members; hash-mode selects IP-pair hashing. The sla subtable attaches Internet-Probe target 1. priority-members 1 2 lists eligible members, not 2:1 weights.
Put this general rule after critical rules. IDs do not establish match order; inspect table order separately. Lowest cost SLA configuration; hash-mode in CLI Reference 7.6.3
6. Default route, firewall policy and NAT
config router static
edit 0
set dst 0.0.0.0 0.0.0.0
set sdwan-zone "internet-zone"
set distance 10
next
end
config firewall policy
edit 0
set name "LAN-to-Internet-SDWAN"
set srcintf "internal"
set dstintf "internet-zone"
set srcaddr "LAN-10.10.10.0_24"
set dstaddr "all"
set action accept
set schedule "always"
set service "ALL"
set nat enable
set logtraffic all
next
end
In these tables, edit 0 creates entries with allocated IDs. dst 0.0.0.0 0.0.0.0 denotes the default route, and sdwan-zone attaches it to the zone; actual next hops come from members. distance 10 is administrative distance, not a distribution weight. Inspect older/automatic lower-distance defaults, specific routes and policy routes. Do not rerun this block repeatedly for tests because it creates entries. Static routing CLI
In the policy, srcintf/dstintf define LAN-to-zone security flow, srcaddr/dstaddr the addresses, action accept permission, schedule always timing, and service ALL the example services. nat enable uses the outgoing interface address for SNAT; logtraffic all logs accepted traffic.
This is a connectivity baseline; apply organizational access restrictions and security profiles before operation. Stateful replies do not require a general WAN-to-LAN policy. Configuring SD-WAN in the CLI
Central NAT: the example assumes it is disabled. If enabled, set nat enable in the policy is insufficient; an appropriate Central SNAT map must match the existing design. Do not blindly disable Central NAT on a live network to fit this example. Central SNAT
Configuration acceptance tests
Before changing anything, create new sessions from multiple clients to multiple destinations and record members, rule IDs and egress IPs. Then test physical WAN1 loss, upstream failure with a healthy gateway, quality degradation and recovery, loss of one probe target, and both WANs outside SLA.
Acceptance: both healthy members are eligible; new connections use WAN2 when WAN1 fails; after recovery, new sessions distribute according to the rule; and one target's failure is not conclusively reported as an ISP outage. Record existing-session behavior, egress IPs, voice and sensitive applications separately.
Security Considerations
SD-WAN is not a security policy. WAN-selection rules do not authorize traffic; firewall policy must grant least required access. service ALL only demonstrates connectivity; organizational policy must restrict allowed services, users and destinations.
Apply IPS, antivirus, web/DNS filtering and Application Control according to needs and hardware capacity. TLS deep inspection requires CA management, approved exceptions and compatibility checks. More Internet capacity does not itself increase inspection capacity.
Do not unnecessarily expose management GUI/SSH on public WANs. Restrict management to defined networks or organizational VPNs with identity controls. SNAT does not replace security policy.
Captures/debug may record sensitive IPs, hostnames and connection details. Use narrow, short captures and controlled retention. Turn debug off afterward; do not globally bypass RPF/state controls to fix asymmetry.
This guide covers outbound Internet. Publishing inbound services through two ISPs needs VIP, DNS, inbound availability and response-path design; outbound balancing does not provide this automatically.
Weighted load balancing for 200 and 100 Mbps WANs
Nominal capacity ratio:
200 : 100 = 2 : 1
WAN1 Weight = 2
WAN2 Weight = 1
This ratio is a starting point. Consider upload capacity, guaranteed throughput, congestion and ISP quality. Session ratios do not equal byte ratios; even 2:1 distribution can saturate WAN2 with a few large downloads.
Valid weighted syntax in the implicit rule
The next block completely changes the implicit algorithm for the main scenario's members; it is neither pseudocode nor weighted hashing for rule 10:
config system sdwan
set load-balance-mode weight-based
config members
edit 1
set weight 2
next
edit 2
set weight 1
next
end
end
load-balance-mode weight-based activates implicit session distribution. weight defines each member's nonzero relative share. This syntax exists in CLI References 7.4.6 and 7.6.3. Explicit General-Internet still uses hash-mode source-dest-ip-based; these weights do not make it 2:1. Because General-Internet matches the LAN, this change will not produce the expected weighted result for that traffic. CLI Reference 7.4.6; CLI Reference 7.6.3
To use implicit weighting for LAN traffic, redesign explicit-rule coverage; then it is no longer the same SLA-aware source-destination architecture. Do not combine implicit weights with explicit hashing and promise undocumented results.
For unequal links, retain general hashing with download/backup steering and QoS, or use an explicit-rule available-bandwidth algorithm while accepting changed persistence. inbandwidth, outbandwidth and bibandwidth exist in Reference 7.6.3; record interface capacity correctly and use actual upload capacity rather than assuming download symmetry. Manual strategy — bandwidth algorithms
Traffic steering: SD-WAN beyond load distribution
| Suggested priority | Traffic | Strategy | Operational note |
|---|---|---|---|
| 1 | VoIP | Best Quality; jitter or a suitable combined profile | Actual media-provider requirements and QoS matter |
| 2 | Microsoft 365 / Business Apps | SLA or Best Quality | Service-specific probes and valid application/ISDB identification |
| 3 | Backup / Large Downloads | Secondary or cheaper link with fallback | Scheduling and shaping to protect sensitive services |
| Last | General Internet | SLA + Source-Destination Hash | General rule after exceptions |
Best Quality using latency does not necessarily minimize jitter too. Choose jitter as the main metric or design a custom profile; “best latency/jitter” needs an operational definition. The general two-DNS SLA is not a Microsoft 365 service SLA.
Working example: steering one voice host
Here, 10.10.10.50 is a dedicated voice gateway whose traffic should follow jitter-based selection. The existing LAN-to-zone policy covers it. This is address-based and does not automatically identify all LAN VoIP.
config firewall address
edit "Voice-Gateway"
set subnet 10.10.10.50 255.255.255.255
next
end
config system sdwan
config service
edit 5
set name "Voice-Best-Jitter"
set mode priority
set src "Voice-Gateway"
set dst "all"
set health-check "Internet-Probe"
set link-cost-factor jitter
set link-cost-threshold 10
set priority-members 1 2
next
move 5 before 10
end
end
The address object matches the voice host. mode priority means Best Quality; health-check supplies measurements; link-cost-factor jitter ranks paths; link-cost-threshold 10 sets the percentage threshold for changing preference. priority-members lists candidates. move 5 before 10 places the specific rule before the general one. In production, replace public probes with a voice-provider-specific check; this only demonstrates steering. Best Quality CLI
Failover scenarios and session behavior
WAN1 alive + SLA pass; WAN2 alive + SLA pass
→ New sessions distributed by source/destination hash
WAN1 dead OR SLA fail; WAN2 alive + SLA pass
→ New sessions use WAN2
WAN1 recovered + SLA pass
→ New sessions follow the configured SD-WAN rule again
In active/active, “WAN1 healthy → normal traffic” does not mean all traffic uses WAN1; WAN2 participates from the start. Recovery does not return every connection to WAN1; eligible paths rejoin rule decisions.
New Sessions
New connections select among usable members according to the rule. In the main rule, an unsuitable WAN1 and healthy WAN2 mean WAN2 for new sessions. Test rule, route, SLA and hash changes with new sessions; reopening a page may reuse existing TCP/HTTP2/QUIC sessions.
Existing Sessions
Existing sessions have paths, TCP/UDP state and NAT. Brownouts may leave them on the previous path; link-down or route changes can trigger reevaluation or disconnection. Exact behavior depends on route changes, session settings, SNAT, protocol and patch.
With dual ISPs, WAN1 SNAT might be 203.0.113.2 and WAN2 198.51.100.2. A destination cannot treat an ordinary TCP connection with a changed public source IP as the same connection. Downloads may need retry/resume, calls may drop and VPNs may reconnect. Seamless migration of all sessions cannot be promised.
Settings such as snat-route-change affect SNAT route reevaluation, but do not make TCP resilient to public-IP changes. Do not enable them in production merely for “better failover”; scope, release behavior and application tests are essential. CLI global setting
Reduce flapping with suitable recovery settings, baseline-derived thresholds and hold-down where the strategy supports it. Two-FortiGate HA alone does not solve public-IP changes between ISPs.
Troubleshooting
Compare diagnostic output with the indicators described in each section. Formatting, interface indexes and flags can vary by model and patch. Confirm the correct VDOM first; a different field label or interface index alone is not a fault.
Figure 5 — Diagnose from interface health through rules, routes, policies, NAT and packet observation; a healthy gateway alone does not prove Internet health.
Troubleshooting flow
Internet Problem
↓
Interface Up? ── No → Cable / modem / admin status / errors
↓ Yes
Gateway Reachable? ── No → IP / mask / ARP / ISP handoff
↓ Yes
Performance SLA Passed? ── No → Probe / upstream / quality
↓ Yes
Correct SD-WAN Rule Matched? ── No → match criteria / order
↓ Yes
Route Exists? ── No → default / distance / route withdrawal
↓ Yes
Firewall Policy Matched? ── No → interfaces / objects / order
↓ Yes
NAT Correct? ── No → outgoing IP / IP pool / central SNAT
↓ Yes
Session / Packet Debug
↓
Return traffic / DNS / MTU / application / offload
This is an investigation flow, not FortiOS packet-processing order. Recheck dependencies from the start after each correction.
1. Version and interface state
Command
get system status
get system interface physical
diagnose hardware deviceinfo nic wan1
diagnose hardware deviceinfo nic wan2
Expected output: firmware version and VDOM mode; interface IP/link status; NIC speed/duplex and RX/TX/error counters. NIC field names depend on the model.
Interpretation: administrative up differs from link up. Resolve link-down before SD-WAN. Increasing errors/drops or bad negotiation can degrade SLA. For a software/hardware switch named internal, physical summaries alone do not fully describe logical state. CLI troubleshooting cheat sheet
2. Members and zone
Command
diagnose sys sdwan member
diagnose sys sdwan zone
show system sdwan
Expected output: member 1 uses wan1/gateway 203.0.113.1; member 2 uses wan2/198.51.100.1; zone internet-zone. show displays saved configuration.
Interpretation: incorrect gateways, other zones or different IDs separate rules/routes from the intended scenario. Configuration display alone does not prove runtime use. SD-WAN diagnostics; diagnose sys reference
3. Gateway and external reachability through each WAN
Command
execute ping-options reset
execute ping-options interface wan1
execute ping-options source 203.0.113.2
execute ping 203.0.113.1
execute ping 1.1.1.1
execute ping-options reset
execute ping-options interface wan2
execute ping-options source 198.51.100.2
execute ping 198.51.100.1
execute ping 8.8.8.8
execute ping-options reset
get system arp
Expected output: destination echo replies, low packet loss and an ARP entry for each WAN gateway. Some ISPs block gateway ping, so missing replies alone do not conclusively prove failure.
Interpretation: a responding gateway with an unresponsive external target suggests upstream trouble or probe filtering. Failed ARP may indicate wrong IP/mask, ISP VLAN or cabling. This FortiGate-originated ping does not replace user forwarding and policy/NAT tests. The final reset prevents forced source/interface settings leaking into later tests. Ping options reference; ARP and network commands
4. Performance SLA
Command
diagnose sys sdwan health-check Internet-Probe
Expected output — teaching example
Health Check(Internet-Probe):
Seq(1 wan1): state(alive), packet-loss(0.000%) latency(42.000), jitter(4.000) sla_map=0x1
Seq(2 wan2): state(alive), packet-loss(0.000%) latency(65.000), jitter(8.000) sla_map=0x1
Interpretation: state indicates responsiveness; metrics indicate quality. With one SLA target ID 1, the relevant pass bit should be set. sla_map is a bitmap; do not interpret multi-target values without mapping. Alive and SLA failed can coexist. Compare thresholds and rule state. Health-check output example
5. SD-WAN rule and selected member
Command
diagnose sys sdwan service4 10
Expected output: rule 10, SLA state and selected members. Runtime mode may display Load-balance although configuration uses mode sla and load-balance enable.
Interpretation: two healthy members should be eligible. If WAN1 fails and WAN2 passes, expect WAN2. This shows rule state, not proof that a particular client matched; inspect that client's actual rule ID in the session table. Official service4 example
6. Routes
Command
get router info routing-table all
get router info routing-table database
show router static
Expected output: active defaults through eligible WAN next hops and a connected 10.10.10.0/24 route. Conceptual healthy default: 0.0.0.0/0 via 203.0.113.1, wan1, plus 198.51.100.1, wan2.
Interpretation: route output may show real members rather than just the zone. Saved routes are not necessarily active. Check other lower-distance routes, more-specific prefixes, policy routing, wrong VRFs and next-hop removal by health checks. Verifying routing table; Routing diagnostics
7. Actual sessions, policy and NAT
Command
diagnose sys session filter clear
diagnose sys session filter src 10.10.10.25
diagnose sys session list
diagnose sys session filter clear
Expected output: client sessions, policy_id, forward/reverse interfaces, NAT and applicable sdwan_mbr_seq/sdwan_service_id. The general SD-WAN rule should be ID 10; the policy ID was allocated by edit 0.
Interpretation: sdwan_mbr_seq=1 or 2 identifies the selected path. NAT must match the outgoing WAN; inspect act=snat and translated tuples in hooks. Do not extrapolate current routes to old sessions. filter clear only clears display filters, not sessions. Session tracking example; Session diagnostics
8. Firewall policy and Central NAT
Command
show firewall policy
show system settings
show firewall central-snat-map
Expected output: internal → internet-zone policy, LAN source, Accept and NAT for policy-based NAT; Central NAT state and matching map if enabled.
Interpretation: a policy must match in actual order. An earlier general deny or narrower policy may match first. With Central NAT enabled, inspect SNAT in its map. An ISP-1 IP pool on WAN2 can create invalid sources and unanswered traffic. Sessions/debug provide final policy/NAT confirmation. Central SNAT behavior
9. Limited packet capture
Command
diagnose sniffer packet any 'host 10.10.10.25 or host 8.8.8.8' 4 50 l
Expected output: ingress packets on internal and egress on the selected WAN, with both directions for ping to 8.8.8.8. Verbosity 4 shows interfaces; capture ends after 50 packets or can be stopped with Ctrl+C when traffic is absent.
Interpretation: after SNAT, WAN packets no longer carry the client's private IP, hence the destination filter too. Unanswered WAN egress suggests ISP/destination/NAT issues; replies on another WAN suggest asymmetry. This filter may include other users' traffic to that destination; minimize capture. Sniffer trace reference
10. Debug Flow for a fresh connection
Command — start
diagnose debug reset
diagnose debug flow filter clear
diagnose debug flow filter saddr 10.10.10.25
diagnose debug flow filter daddr 8.8.8.8
diagnose debug flow show function-name enable
diagnose debug flow trace start 50
diagnose debug enable
Now create a fresh connection from the client. Ping suits path checks; HTTPS requires a real service destination and its corresponding filter.
Command — stop and clean up
diagnose debug disable
diagnose debug flow trace stop
diagnose debug flow filter clear
diagnose debug reset
Expected output: session allocation, selected routes, policy matches, SNAT or drop reasons. Messages such as Allowed by Policy or Denied by forward policy check vary by outcome/build.
Interpretation: distinguish missing routes, policy denies, RPF/asymmetry and invalid SNAT. Old or hardware-offloaded flows may yield little debug output. Create a new session first; disable offload only for a limited investigation after load assessment. Ordinary Debug Flow shows CPU processing only. Debug Flow CLI; NPU diagnostics; GUI Debug Flow filters
When IP works but websites do not
After routes and policies pass, inspect DNS, MTU/PMTUD, MSS, TLS inspection, proxies and destination restrictions. Small pings do not prove large HTTPS transfers work. Public-target ping does not prove the user's resolver works. Look for SYN/SYN-ACK, retransmissions, ICMP Fragmentation Needed and resets.
Do not clear the whole session table to troubleshoot; this can disrupt calls, VPNs and users. New connections and narrow filters are usually sufficient for selection tests.
Common Mistakes
| Mistake | Effect | Correction |
|---|---|---|
| Default route to wrong WAN or gateway | Bypasses intended selection or blackholes traffic | Verify the zone and member gateways |
| Overlapping static routes | Unexpected selection by distance/prefix | Inspect active routes, the database and policy routes |
| Incorrect NAT | Packets leave with private or other-ISP addresses | Match SNAT to outgoing WAN; check Central NAT/IP pools |
| Missing firewall policy | Drops despite healthy routes and SLAs | Policy between LAN and SD-WAN zone |
| Probing only the ISP gateway | Misses upstream failures | External targets and service checks |
| Probing only one public IP | Detection depends on one destination | Fallback servers and independent checks |
| Unsuitable SLA thresholds | Flapping or missed degradation | Service baselines and SLOs |
| Expecting persistence across an entire website | Different IPs across one application's destinations | Dedicated sensitive-application rule |
| Asymmetric Routing | RPF/state problems | Design return paths and NAT; fix the cause |
| Wrong SD-WAN rule priority | Catch-all matches before a specific rule | Actual rule order, not just IDs |
| 50/50 distribution on 200/100 WANs | WAN2 may saturate | Steering, QoS or capacity-appropriate methods |
| Member weights alongside explicit hashing | Incorrect expectation of 2:1 distribution | Separate implicit algorithms from explicit hashes |
| Treating SLA failure as dead | Misinterprets routes and failover | Observe state, SLA bitmap and rule runtime |
| Checking only old sessions | Incorrect conclusions about new rules | Create fresh connections |
Equal hash/session counts on unequal links are not automatically defects; the mistake is calling that optimal bandwidth distribution without capacity analysis.
Best Practices
- Choose physically and upstream-independent ISPs; two contracts do not necessarily provide two failure domains.
- Supplement public probes with independent critical-service checks whose protocols fit the operational question.
- Give critical applications separate rules instead of aggregating all traffic under one default.
- Put general source-destination rules after specific rules and document final fallback.
- Record actual upload/download capacity and peaks; equal hashes need not mean equal Mbps.
- Size backup WAN capacity for essential load and throttle backups/updates during outages.
- Record quality baselines, failover timing, SLA failures and recovery; alert on meaningful changes.
- Test cable loss, healthy-gateway/upstream failure, brownouts, one failed target and both failed WANs.
- Test IP-allowlisted applications, SSO, banking and SIP separately, pinning them where necessary.
- Manage routes, firewall policies, NAT, SD-WAN rules and monitoring as one reversible change.
- Select model-appropriate firmware using release notes and organizational maintenance policy; the reference version is not an automatic upgrade recommendation.
- Evaluate algorithm, rule-order and SLA changes with fresh sessions and real traffic.
Conclusion: proposed production architecture
Evaluate the balancing algorithm alongside path quality, application policy, connection stability and link capacity before handover. Use source-destination hashing in an explicit SLA rule for general internet access and route critical applications separately. Record both new-connection success and existing-session limitations during failover testing so the team has clear expectations for the two WANs.
FortiGate SD-WAN
├── internet-zone
│ ├── WAN1 — ISP-1 — 200 Mbps
│ └── WAN2 — ISP-2 — 100 Mbps
│
├── Performance SLA
│ ├── Independent / application-relevant probes
│ ├── Latency
│ ├── Jitter
│ └── Packet Loss
│
├── Critical Applications
│ └── Best Quality / service-specific SLA / defined fallback
│
├── Backup / Large Downloads
│ └── Secondary or lower-cost WAN + shaping
│
└── General Internet
└── SLA-aware Source-Destination Load Balancing
└── Automatic failover for new sessions
Production readiness means valid syntax, explicit assumptions, documented fallback, correct policy/NAT and actual-device acceptance tests. This architecture increases aggregate capacity across multiple connections; it does not turn one TCP session into a 300 Mbps link or promise impact-free failover for every session.
FAQ
Can FortiGate use two Internet links simultaneously?
Yes. SD-WAN with a load-balancing rule lets different connections use both eligible WANs. Sensitive traffic can use dedicated rules.
Are the two Internet speeds added together?
Many independent connections can use aggregate capacity from both links; an ordinary session does not get their combined speed. Practical limits depend on load, destinations, quality, NAT and inspection capacity.
Can one download use both WANs simultaneously?
Not one normal TCP session. Multi-connection downloads may use several WANs with some methods; source-destination hashing usually keeps the same IP pair on one WAN.
How do source-IP and source-destination-IP methods differ?
Source-IP maps a client's destinations to one path. Source-destination maps each pair separately, allowing one client to use different WANs for different destinations.
What happens to sessions if the primary WAN fails?
New connections use a healthy path. Existing connections may fail, retry or reconnect; changed public SNAT addresses prevent a promise of uninterrupted TCP migration.
How does Performance SLA differ from a simple ping?
Simple ping usually checks immediate reachability. SLA adds continuous metric measurement, thresholds and WAN-selection integration. Health checks can use ping or another service-appropriate protocol.
Which method suits two links with different speeds?
For 200/100 Mbps, roughly 2:1 session weights can be a starting point, but implicit weighting and explicit SLA rules differ. For general persistence, combine hashing with steering/QoS; for capacity-based distribution, test a method compatible with the rule and release.
Is ECMP or SD-WAN better?
ECMP suits simple multipath routing. SD-WAN provides fuller policy tools when quality, applications, fallback and monitoring matter. SD-WAN still relies on valid routes and routing components; the concepts are related.
Does SLA failure always remove the default route?
No. Alive/dead differs from quality pass/fail. An alive, slow path may retain its route while a rule sends new traffic through another WAN.
What if both WANs fall outside SLA?
If both remain alive, this strategy's general rule may use both to preserve connectivity. Design and test fail-closed or other requirements separately.
Do two health-check probes mean simultaneous testing and majority voting?
No. This server model uses the first server, then the second when the first is unavailable. Independent destination monitoring needs separate health checks and clear rule logic.
Is session-based the same as packet-based load balancing?
No. Distributing new sessions does not split every packet of one TCP flow between ISPs. Ordinary session packets need compatible paths and state.
References
Use the official sources below for the behavior and syntax discussed in each section. When adapting the example, reassess link capacity, session patterns and service policies for your network.
- FortiOS 7.4.6 — SD-WAN Rules
- FortiOS 7.4.1 — Selecting the implicit SD-WAN algorithm
- FortiOS 7.4.1 — Implicit rule
- FortiOS 7.4.7 — Load balancing strategy
- FortiOS 7.4.6 — config system sdwan
- FortiOS 7.6.0 — config system sdwan
- FortiOS 7.6.3 — config system sdwan
- FortiOS 7.6.0 — Lowest cost SLA
- FortiOS 7.6.4 — Lowest cost SLA and fallback
- FortiOS 7.6.4 — Link health monitor
- FortiOS 7.6.3 — Configuring SD-WAN in the CLI
- FortiOS 7.6.0 — config router static
- FortiOS 7.6.3 — Central SNAT
- FortiOS 7.6.3 — Best quality strategy
- FortiOS 7.6.3 — Manual strategy and available bandwidth
- FortiOS 7.6.1 — Dynamic application steering
- FortiOS 7.6.3 — config system global
- FortiOS 7.6.0 — CLI troubleshooting cheat sheet
- FortiOS 7.6.3 — SD-WAN related diagnose commands
- FortiOS 7.6.3 — diagnose sys reference
- FortiOS 7.6.6 — Session diagnostics reference
- FortiOS 7.6.0 — SD-WAN configuration and session fields
- FortiOS 7.6.0 — Diagnosing NPU-based interfaces
- FortiOS 7.6.3 — Using the debug flow tool
- FortiOS 7.6.3 — Interface configuration example
- FortiOS 7.6.3 — Address object example
- FortiOS 7.6.0 — Interface CLI reference
- FortiOS 7.6.6 — Ping options reference
- FortiOS 7.4.8 — Verifying routing table
- FortiOS 7.6.4 — Sniffer trace reference
- FortiOS 7.6.6 — Debugging packet flow
- FortiOS 7.6.3 — get system status example
- FortiOS — CLI table subcommands