BGP Tuning

! Set global TCP MSS and enable PTMUD
 ip tcp mss XXXX
 ip tcp path-mtu-discovery ! Enabled by default


interface X/X
 mtu 9192           ! Increase the MTU on an eBGP facing link to reduce the load time for a full
                    ! table peer, by reducing the TCP overhead
 hold-queue xxx in  ! We can increase the input hold-queue for packets that are being CPU punted 
                    ! (such as BGP packets) to avoid drops and TCP retransmisions

router bgp 65000

 ! Enable BGP per peer PMTUD for a specific peer
 neighbor x.x.x.x transport path-mtu-discovery

 ! Enable BGP PMTUD for all peers, this is enabled by default
 bgp transport path-mtu-discovery

 ! We don't really want to disable this, as the event is driven by an interface going down or eBGP neighbour next-hop for multihop BGP going down, although if we have a flappy IXP interface we might?  This is already disabled for iBGP neighbours by default.
 no bgp fast-external-fallover

 ! BGP Next Hop Tracking (NHT) is enabled by default, so apart from disabling it we can lower the trigger delay, which is default 5 seconds and not really worth reducing, only possible increasing on older platforms with many VPNv4 prefixes perhaps?
 ! Consider IGP convergence time in relation to this - rapid convergence means many BGP best-path process runs
 bgp nexthop trigger delay 5
show ip bgp attr nexthop rib ! Hidden command

 ! Default is 60 seconds for main RIB and 15 for VPNV4 RIB
 bgp scan-time 60 

 ! Think about raising the update delay timer if we have many full table peers or if we are a route reflector client (default 120 seconds).
 ! For slower platforms like 7600s with full table feeds, we might want to extend the read-only period at BGP start up
 bgp update-delay 180

 ! Updates are sent to iBGP neighbours every 5 seconds (but they should already be up to date via the IGP) and eBGP neighbours every 360 seconds. This could be lowered on agreed eBGP neighbours (confederations perhaps / inter-AS BGP signalled MPLS perhaps ?) or iBGP neighbours that are VPNv4 PEs perhaps?
 neighbor x.x.x.x advertisement-interval 0

 ! BFD is preferred neighbour detection loss method, use instead of fast-externa-failover above
 neighbor x.x.x.x fall-over bfd

 ! Lower the BGP hold timer if the external AS agrees, within the IGP for iBGP neighbours we rely on OSPF/ISIS or BFD for link failure (default 60 and 180 seconds)
 neighbor x.x.x.x timers 5 15

---