Chapter 9 High Availability for FortiOS 5.0 : HA and load balancing : Load balancing overview : Using FortiGate network processor interfaces to accelerate active-active HA performance
  
Using FortiGate network processor interfaces to accelerate active-active HA performance
Many FortiGate models and FortiGate AMC modules include network processors that can provide hardware acceleration for active-active HA load balancing by offloading load balancing from the primary unit CPU. HA load balancing can be accelerated by interfaces accelerated by NP network processors.
In some cases, performance of the primary unit can be reduced by active-active HA load balancing. Primary unit CPU cycles and bus bandwidth are required to receive, calculate load balancing schedules, and send balanced packets to the subordinate units. In very busy active-active clusters the primary unit may not be able to keep up with the processing load. This can result in lost traffic and can also cause the primary unit to delay sending heartbeat packets possibly reducing the stability and reliability of the active-active HA cluster.
Adding network processors to busy cluster unit interfaces increases load balancing performance by offloading load balancing to the network processors. The first packet of every new session is received by the primary unit and the primary unit uses its load balancing schedule to select the cluster unit that will process the new session. This information is passed back to the network processor and all subsequent packets of the same sessions are received by the primary unit interface network processor which sends the packet directly to a subordinate unit without using the primary unit CPU. Load balancing is effectively offloaded from the primary unit to the network processor resulting in a faster and more stable active-active cluster.
Using network processors to accelerate load balancing is especially useful if the load-balance-all and load-balance-udp options are enabled and the cluster is load balancing all TCP and UDP sessions because this could mean that the cluster is load balancing an excessive number of sessions.
To take advantage of network processor load balancing acceleration, connect the cluster unit interfaces with network processors to the busiest networks. Connect non‑accelerated interfaces to less busy networks. No special FortiOS or HA configuration is required. Network processor acceleration of active-active HA load balancing is supported for any active‑active HA configuration or active-active HA load balancing schedule.