BGP peer sessions in Active state causing high CPU load

[reinier.hamstra]

Hi all,

Does anybody have any experience with BGP peering sessions in Active state causing high CPU load?

We are currently running two 7604s with Sup720-3BXL. Main router is currently switching around 6gigs in traffic, with two peering sessions each providing full routes. Additionally there are around 100 operational peering sessions at a local exchange. No complex configurations, these machines do peering only.

Average cpu load was 30-35% until I administratively shut about 15 inactive sessions that were in Active state. Since then CPU load has dropped to around 17%?!

I did expect some performance gain, yet this effect seems quite disproportionate. I have not been able to find any satisfactory answer why Active state sessions would consume so many resources.

Any thoughts or experiences are very welcome!

[umedryk]

While they help simplify BGP configuration, BGP peer groups also can enhance scalability. All peer group members must share a common outbound policy. Thus, the same update packets can be sent to each group member, reducing the number of CPU cycles that BGP requires to advertise routes to peers. In other words, with peer groups, BGP walks the BGP table only on the peer group leader, filters the prefixes through the outbound policies, and generates updates, which it sends to the peer group leader. In turn, the leader replicates the updates to group members with which it is synchronized. Without peer groups, BGP must walk the table for every peer, filter prefixes through outbound policies, and generate updates that are sent only to the one peer.

[Mohamed Sobair]

Hi,

Most of the cases that causes a bgp session to be in active state are as follows:

1-Make sure that you peer with the correct neighbor address specially when using the (Update-source Loopback) statment, and vice versa for IBGP and EBGP peering, when this is configured the peering router should peer with your loopback address.

2- For EBGP neighbors, if they are not directly connected, makes sure you configure (EBGP multi-hop TTL) neighbor command, because by default EBGP neighbors have a TTL of 1 , and could be overrided by (EBGP multi-hop TTL) command.

HTH

Mohamed Sobair

[Edison Ortiz]

Please see:

http://www.cisco.com/warp/public/459/highcpu-bgp.html

HTH,