BGP Performance Tuning Overview

Introduction

BGP performance tuning has many different layers that apply to a number of cases. Some of these apply to cold starts of BGP while others deal with changes on a network such as failures. This page will list some of the tuning features related to BGP's convergence.

BGP Performance Tuning

Most of these feature relate to optimizing the full loading of the routing table.

TCP Tuning

  • TCP MSS - Some IOS version could have the MSS set to 536. Recommended to set it to MTU - IP/TCP headers using PMTUD.
  • Path MTU Discovery - PMTU will discover the MTU of a path to set the correct MSS.

Router Queues

Sizing the router queues can be important for large BGP updates with many peers. Processing all of the TCP ACK can be difficult with default settings.

  • Input Hold Queue - Eliminating input hold drops increases the BGP performance. Usually this has to do with the very large number of TCP ACKs being dropped.
  • Selective Packet Discard (SPD) - Control Plane Policing mechanism which protects traffic towards the processor.
  • System Buffers - Used for packet processing and generations of packets to the processor, once they are taken form input hold queue of an interfaces. Usually this relates to BGP by increasing the number of small buffers to be able to process the ACKs.

BGP Update Optimization

  • Peer Groups - Optimize the number of BGP Update messages that are generated. Instead of generating the same Update message to all of its peers, messages are replicated based on one peer. These are configured manually.
  • Dynamic Update Peer Groups - Dynamically created update peer groups based on the same BGP policy. Does not require manual configuration and performs the same optimization of BGP Update message replication.
  • BGP Read-Only Mode - Waits for the Best Path selection until all updates are received by getting BGP Keepalive or bgp update-delay timer expires. Keepalive is an indication that the initial routing updates are done. It prevents routers from constantly sending updates for paths that might not have all of the available paths.
  • Selective Route Download - Applies to out of band RR, with large number of reflected routes that don't need to be installed in the FIB and RIB (out of band).
  • Slow Peer Detection/Protection

BGP Network Convergence

These features relate to the operation BGP and any changes that occur after the BGP table is fully converged from initial state.

  • BGP Fast External Failover - When an interface to an eBGP peer goes down, fast external failover will remove all prefixes from that peer and drop the BGP session, without waiting for Holdtimer to expire.
  • IGP/BGP Convergence - When a router reboots or is brought back to services, usually IGP converges faster than BGP, calculating a best path through that device can cause blackholing of traffic, since not all BGP peers might have converged to the same NEXT_HOP. To prevent this OSPF process uses MAX Metric and ISIS uses the Overload-bit. Both of these have an option to wait for BGP to converge or set timer after reboot to advertise an LSP that is not preferred.
  • BGP Non-Stop Forwarding (NSF) - When a device with redundant RPs, reboots one of the RP, NSF is able to maintain the data plane forwarding while the redundant RP comes online. Requires BGP protocol extension.
  • BGP Non-Stop Routing (NSR) - Similar to NSF, but does not require protocol extensions. The redundant RP synchronize all of their BGP data and take over the TCP session during a failover.
  • BGP Prefix Independent Convergence
  • BGP Additional-Path
  • BGP Best-External

Prefix Update Optimization

  • Route Dampening
  • Soft-Reconfiguration
  • Router Refresh
  • Transmit Side AS-Loop Detection
  • Outbound Route Filtering (ORF)

Additional Resources

BRKRST-3321 - Scaling BGP by Luc De Ghein Cisco Live Video
BRKRST-3371 - Advances in BGP by Gunter Van de Velde Cisco Live Video

Comments

rating: 0+x
Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-ShareAlike 3.0 License