Mitigating Negative Impact of Maintenance through BGP Session Culling
LONAP Ltd
5 Fleet Place
London
EC4M 7RD
United Kingdom
will@lonap.net
20C
1658 Milwaukee Ave # 100-4506
Chicago
IL
60647
United States of America
grizz@20c.com
NTT Communications
Theodorus Majofskistraat 100
1065 SZ
Amsterdam
The Netherlands
job@ntt.net
INEX
4027 Kingswood Road
Dublin
24
Ireland
nick@inex.ie
Routing
Global Routing Operations
BGP
culling
EBGP
sessions
This document outlines an approach to mitigate negative impact on networks resulting from maintenance activities.
It includes guidance for both IP networks and Internet Exchange Points (IXPs).
The approach is to ensure BGP-4 sessions affected by the maintenance are forcefully torn down before the actual maintenance activities commence.
BGP Session Culling is the practice of ensuring BGP sessions are forcefully torn down before maintenance activities on a lower layer network commence, which otherwise would affect the flow of data between the BGP speakers.
BGP Session Culling ensures that lower layer network maintenance activities cause the minimum possible amount of disruption, by causing BGP speakers to preemptively gracefully converge onto alternative paths while the lower layer network's forwarding plane remains fully operational.
The grace period required for a successful application of BGP Session Culling is the sum of the time needed to detect the loss of the BGP session, plus the time required for the BGP speaker to converge onto alternative paths.
The first value is governed by the BGP Hold Timer (section 6.5 of ), commonly between 90 and 180 seconds,
The second value is implementation specific, but could be as much as 15 minutes when a router with a slow control-plane is receiving a full set of Internet routes.
Throughout this document the "Caretaker" is defined to be the operator of the lower layer network, while "Operators" directly administrate the BGP speakers.
Operators and Caretakers implementing BGP Session Culling are encouraged to avoid using a fixed grace period, but instead monitor forwarding plane activity while the culling is taking place and consider it complete once traffic levels have dropped to a minimum.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.
From the viewpoint of the IP network operator, there are two types of BGP Session Culling:
The operator initiates the tear down of the potentially affected BGP session by issuing an Administrative Shutdown.
The caretaker of the lower layer network disrupts BGP control-plane traffic in the upper layer, causing the BGP Hold Timers of the affected BGP session to expire, subsequently triggering rerouting of end user traffic.
Before an operator commences activities which can cause disruption to the flow of data through the lower layer network, an operator can reduce loss of traffic by issuing an Administratively Shutdown to all BGP sessions running across the lower layer network and wait a few minutes for data-plane traffic to subside.
While architectures exist to facilitate quick network reconvergence (such as BGP PIC), an operator cannot assume the remote side has such capabilities.
As such, a grace period between the Administrative Shutdown and the impacting maintenance activities is warranted.
After the maintenance activities have concluded, the operator is expected to restore the BGP sessions to their original Administrative state.
Initiators of the Administrative Shutdown could consider to use [Graceful Shutdown] to facilitate smooth drainage of traffic prior to session tear down, and the Shutdown Communication to inform the remote side on the nature and duration of the maintenance activities.
In the case where multilateral interconnection between BGP speakers is facilitated through a switched layer-2 fabric, such as commonly seen at Internet Exchange Points (IXPs), different operational considerations can apply.
Operational experience shows many network operators are unable to carry out the Voluntary BGP Session Teardown recommendations, because of the operational cost and risk of co-ordinating the two configuration changes required. This has an adverse affect on Internet performance.
In the absence of notifications from the lower layer (e.g. ethernet link down) consistent with the planned maintenance activities in a densely meshed multi-node layer-2 fabric, the caretaker of the fabric could opt to cull BGP sessions on behalf of the stakeholders connected to the fabric.
Such culling of control-plane traffic will pre-empt the loss of end-user traffic, by causing the expiration of BGP Hold Timers ahead of the moment where the expiration would occur without intervention from the fabric's caretaker.
In this scenario, BGP Session Culling is accomplished through the application of a combined layer-3 and layer-4 packet filter deployed in the switched fabric itself.
The following considerations apply to the packet filter design:
The packet filter MUST only affect BGP traffic specific to the layer-2 fabric, i.e. forming part of the control plane of the system described, rather than multihop BGP traffic which merely transits
The packet filter MUST only affect BGP, i.e. TCP/179
The packet filter SHOULD make provision for the bidirectional nature of BGP, i.e. that sessions may be established in either direction
The packet filter MUST affect all relevant AFIs
contains examples of correct packet filters for various platforms.
Not all hardware is capable of deploying layer 3 / layer 4 filters on layer 2 ports, and even on platforms which support the feature, documented limitations may exist or hardware resource allocation failures may occur during filter deployment which may cause unexpected results.
These problems may include:
Platform inability to apply layer 3/4 filters on ports which already have layer 2 filters applied
Layer 3/4 filters supported for IPv4 but not for IPv6
Layer 3/4 filters supported on physical ports, but not on 802.3ad Link Aggregate ports
Failure of the operator to apply filters to all 802.3ad Link Aggregate ports
Limitations in ACL hardware mechanisms causing filters not to be applied
Fragmentation of ACL lookup memory causing transient ACL application problems which are resolved after ACL removal / reapplication
Temporary service loss during hardware programming
Reduction in hardware ACL capacity if the platform enables lossless ACL application
It is advisable for the operator to be aware of the limitations of their hardware, and to thoroughly test all complicated configurations in advance to ensure that problems don't occur during production deployments.
The caretaker of the lower layer can monitor data-plane traffic (e.g. interface counters) and carry out the maintenance without impact to traffic once session culling is complete.
It is recommended that the packet filters are only deployed for the duration of the maintenance and immediately removed after the maintenance.
To prevent unnecessarily troubleshooting, it is RECOMMENDED that caretakers notify the affected operators before the maintenance takes place, and make it explicit that the Involuntary BGP Session Culling methodology will be applied.
The authors would like to thank the following people for their contributions to this document: Saku Ytti, Greg Hankins, James Bensley, Wolfgang Tremmel, Daniel Rösen, Bruno Decraene, and Tore Anderson.
There are no security considerations.
This document has no actions for IANA.
Example packet filters for "Involuntary BGP Session Teardown" at an IXP with LAN prefixes 192.0.2.0/24 and 2001:db8:2::/64.
A repository of configuration examples for a number of assorted platforms can be found at github.com/bgp/bgp-session-culling-config-examples.