| Internet-Draft | BGP SPF Applicability | February 2024 |
| Patel, et al. | Expires 15 August 2024 | [Page] |
This document discusses the usage and applicability of Link State Vector Routing (LSVR) extensions in data center networks utilizing CLOS or Fat-Tree topologies. The document is intended to provide a simplified guide for the deployment of LSVR extensions.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 15 August 2024.¶
Copyright (c) 2024 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
This document complements [I-D.ietf-lsvr-bgp-spf] by discussing the applicability of the technology in a simple and fairly common deployment scenario, which is described in Section 4.¶
After describing the deployment scenario, Section 5 will describe the reasons for BGP modifications for such deployments.¶
Once the control plane routing protocol requirements are described, Section 6 will cover the LSVR protocol enhancements to BGP to meet these requirements and their applicability to Data Center CLOS networks.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
This document assumes knowledge of existing data center networks and data center network topologies [CLOS]. This document also assumes knowledge of data center routing protocols like BGP [RFC4271], BGP-SPF [I-D.ietf-lsvr-bgp-spf], OSPF [RFC2328], as well as, data center OAM protocols like LLDP [RFC4957] and BFD [RFC5580].¶
Within a Data Center, servers are commonly interconnected the CLOS topology [CLOS]. The CLOS topology is fully non-blocking and the topology is realized using Equal Cost Multi-Path (ECMP). In a CLOS topology, the minimum number of parallel paths between two servers is determined by the width of a tier-1 stage as shown in the figure 1.¶
The following example illustrates multi-stage CLOS topology.¶
Tier-1
+-----+
|NODE |
+->| 12 |--+
| +-----+ |
Tier-2 | | Tier-2
+-----+ | +-----+ | +-----+
+------------>|NODE |--+->|NODE |--+--|NODE |-------------+
| +-----| 9 |--+ | 10 | +--| 11 |-----+ |
| | +-----+ +-----+ +-----+ | |
| | | |
| | +-----+ +-----+ +-----+ | |
| +-----+---->|NODE |--+ |NODE | +--|NODE |-----+-----+ |
| | | +---| 6 |--+->| 7 |--+--| 8 |---+ | | |
| | | | +-----+ | +-----+ | +-----+ | | | |
| | | | | | | | | |
+-----+ +-----+ | +-----+ | +-----+ +-----+
|NODE | |NODE | Tier-3 +->|NODE |--+ Tier-3 |NODE | |NODE |
| 1 | | 2 | | 3 | | 4 | | 5 |
+-----+ +-----+ +-----+ +-----+ +-----+
| | | | | | | |
A O B O <- Servers -> Z O O O
In order to simplify layer-3 routing and operations [RFC7938], many data centers use BGP as a routing protocol to create both an underlay and overlay network for their CLOS Topologies. However, BGP is a path-vector routing protocol. Since it does not create a fabric topology, it uses hop-by-hop EBGP peering to facilitate hop-by-hop routing to create the underlay network and to resolve any overlay next hops. The hop-by-hop BGP peering paradigm imposes several restrictions within a CLOS. It severely prohibits a deployment of Route Reflectors/Route Controllers as the EBGP sessions are congruent with the data path. The BGP best-path algorithm is prefix-based and it prevents announcements of prefixes to other BGP speakers until the best-path decision process has been performed for the prefix at each intermediate hop. These restrictions significantly delay the overall convergence of the underlay network within a CLOS network.¶
The LSVR SPF modifications allow BGP to overcome these limitations. Furthermore, using the BGP-LS NLRI format [RFC7752] allows the LSVR data to be advertised for nodes, links, and prefixes in the BGP routing domain and used for SPF computations.¶
With the BGP SPF extensions [I-D.ietf-lsvr-bgp-spf], the BGP best-path computation and route computation are replaced with OSPF-like algorithms [RFC2328] both to determine whether an BGP-LS SPF NLRI has changed and needs to be re-advertised and to compute the BGP routes. These modifications will significantly improve convergence of the underlay while affording the operational benefits of a single routing protocol [RFC7938].¶
Data center controllers typically require visibility to the BGP topology to compute traffic-engineered paths. These controllers learn the topology and other relevant information via the BGP-LS address family [RFC7752] which is totally independent of the underlay address families (usually IPv4/IPv6 unicast). Furthermore, in traditional BGP underlays, all the BGP routers will need to advertise their BGP-LS information independently. With the BGP SPF extensions, controllers can learn the topology using the same BGP advertisements used to compute the underlay routes. Furthermore, these data center controllers can avail the convergence advantages of the BGP SPF extensions. The placement of controllers can be outside of the forwarding path or within the forwarding path.¶
Alternatively, as each and every router in the BGP SPF domain will have a complete view of the topology, the operator can also choose to configure BGP sessions in hop-by-hop peering model described in [RFC7938] along with BFD [RFC5580]. In doing so, while the hop-by-hop peering model lacks the inherent benefits of the controller-based model, BGP updates need not be serialized by BGP best-path algorithm in either of these models. This helps overall network convergence.¶
The BGP SPF extensions [I-D.ietf-lsvr-bgp-spf] define a new BGP-LS SPF SAFI for announcement of BGP SPF link-state. The NLRI format and its associated attributes follow the format of BGP-LS for node, link, and prefix announcements. Whether the peering model within a CLOS follows hop-by-hop peering described in [RFC7938] or any controller-based or route-reflector peering, an operator can exchange BGP SPF SAFI routes over the BGP peering by simply configuring BGP SPF SAFI between the necessary BGP speakers.¶
The BGP-LS SPF SAFI can also co-exist with BGP IP Unicast SAFI which could exchange overlapping IP routes. The routes received by these SAFIs are evaluated, stored, and announced independently according to the rules of [RFC4760]. The tie-breaking of route installation is a matter of the local policies and preferences of the network operator.¶
Finally, as the BGP SPF peering is done following the procedures described in [RFC4271], all the existing transport security mechanisms including [RFC5925] are available for the BGP-LS SPF SAFI.¶
Normally, the BGP-LS AFI/SAFI is used solely to compute the underlay and is given preference over other AFI/SAFIs. Other BGP SAFIs, e.g., IPv6/IPv6 Unicast VPN would use the BGP-SPF computed routes for next hop resolution. However, if BGP-LS NLRI is also being advertised for controller consumption, there is no need to replicate the Node, Link, and Prefix NLRI in BGP-NLRI. Rather, additional NLRI attributes can be advertised in the BGP-LS SPF AFI/SAFI as required (e.g., BGP-LS TE metric extensions [RFC8571] and BGP-LS segment routing extensions [RFC9085]).¶
As previously stated, BGP SPF can be deployed using the existing peering model where there is a single-hop BGP session on each and every link in the data center fabric [RFC7938]. This provides for both the advertisement of routes and the determination of link and neighboring switch availability. With BGP SPF, the underlay will converge faster due to changes to the decision process that will allow NLRI changes to be advertised faster after detecting a change.¶
Alternately, BFD [RFC5580] can be used to swiftly determine the availability of links and the BGP peering model can be significantly sparser than the data center fabric. BGP SPF sessions only need to be established with enough peers to provide a bi-connected graph. If IEBGP is used, then the BGP routers at tier N-1 will act as route-reflectors for the routers at tier N.¶
The obvious usage of sparse peering is to avoid parallel BGP sessions on links between the same two switches in the data center fabric. However, this use case is not very useful since parallel layer-3 links between the same two BGP routers are rare in CLOS or Fat-Tree topologies. Additionally, when there are multiple links, they are often aggregated at the link layer rather than the IP layer. Two more interesting scenarios are described below.¶
In current data center topologies, there is often a very dense mesh of links between levels, e.g., leaf and spine, providing 32-way, 64-way, or more Equal-Cost Multi-Path (ECMP) paths. In these topologies, it is desirable not to have a BGP session on every link and techniques such as the one described in Section 6.2.2 can be used establish sessions on some subset of northbound links. For example, in a Spine-Leaf topology, each leaf switch would only peer with a subset of the spines dependent on the flooding redundancy required to be reasonably certain that every node within the BGP-LS SPF routing domain has the complete topology (refer to Section 6.2.1).¶
Alternately, controller-based data center topologies are envisioned where BGP speakers within the data center only establish BGP sessions with two or more controllers. In these topologies, fabric nodes below the first tier (using [RFC7938] hierarchy) will establish BGP multi-hop sessions with the controllers. For the multi-hop sessions, determining the route to the controllers without depending on BGP would need to be through some other means beyond the scope of this document. However, the BGP discovery mechanisms described in Section 6.5 would be one possibility.¶
With this heuristic, discovery of BGP peers is assumed, e.g., as described in Section 6.5. Additionally, it assumed that the direction of the peering can be ascertained. In the context of a data center fabric, direction is either northbound (toward the spine), southbound (toward the Top-Of-Rack (TOR) switches) or east-west (same level in hierarchy. The determination of the direction is beyond the scope of this document. However, it would be reasonable to assume a technique where the TOR switches can be identified and the number of hops to the TOR is used to determine the direction.¶
In this heuristic, BGP speakers allow passive session establishment for southbound BGP sessions. For northbound sessions, BGP speakers will attempt to maintain two northbound BGP sessions with different switches (in data center fabrics there is normally a single layer-3 connection anyway). For east-west sessions, passive BGP session establishment is allowed. However, BGP speaker will never actively establish an east-west BGP session unless it cannot establish two northbound BGP sessions.¶
One of the advantages of using BGP SPF as the underlay protocol is that BGP policy can be applied at any level. For example, depending upon the topology, it may be possible to aggregate prefix advertisements using existing BGP policy. In Spine/Leaf topologies, it is not necessary to advertise BGP-LS NLRI received by leaves northbound to the spine nodes at the level above. If a common AS is used for the spine nodes, this can easily be accomplished with EBGP and a simple policy to filter advertisements from the leaves to the spine if the first AS in the AS path is the spine AS.¶
In the figure below, the leaves would not advertise any NLRI with AS 64512 as the first AS in the AS path.¶
+--------+ +--------+ +--------+
AS 64512 | | | | | |
for Spine | Spine 1+----+ Spine 2+- ......... -+ Spine N|
Nodes at | | | | | |
this Level +-+-+-+-++ ++-+-+-+-+ +-+-+-+-++
+------+ | | | | | | | | | | |
| +-----|-|-|------+ | | | | | | |
| | +--|-|-|--------+-|-|-----------------+ | | |
| | | | | | +---+ | | | | |
| | | | | | | +--|-|-------------------+ | |
| | | | | | | | | | +------+ +----+
| | | | | | | | | +--------------|----------+ |
| | | | | | | | +-------------+ | | |
| | | | | +----|--|----------------|--|--------+ | |
| | | | +------|--|--------------+ | | | | |
| | | +------+ | | | | | | | |
++--+--++ +-+-+--++ ++-+--+-+ ++-+--+-+
| Leaf 1|~~~~~~| Leaf 2| ........ | Leaf X| | Leaf Y|
+-------+ +-------+ +-------+ +-------+
The most basic requirement is to be able to discover the address of a single-hop peer without pre-configuration. This is being accomplished today with using IPv6 Router Advertisements (RA) [RFC4861] and assuming that a BGP sessions is desired with any discovered peer. Beyond the basic requirement, it is useful to have to following information relating to the BGP session:¶
In a data center fabric, it is often useful to know whether a peer is southbound (towards the servers) or northbound (towards the spine or super-spine), e.g., Section 6.2.2. A potential requirement would be to determine this dynamically. One mechanism, without specifying all the details, might be for the ToR switches to be identified when installed and for the others switches in the fabric to determine their level based on the distance from the closest ToR switch.¶
If there are multiple links between BGP speakers or the links between BGP speakers are unnumbered, it is also useful to be able to establish multi-hop sessions using the loopback addresses. This will often require the discovery protocol to install route(s) toward the potential peer loopback addresses prior to BGP session establishment.¶
Finally, a simple BGP discovery protocol could also be used to establish a multi-hop session with one or more controllers by advertising connectivity to one or more controllers. However, once the multi-hop session actually traverses multiple nodes, it is bordering a distance-vector routing protocol and possibly this is not a good requirement for the discovery protocol.¶
While BGP peer discovery is not part of [I-D.ietf-lsvr-bgp-spf], there are, at least, three proposals for BGP peer discovery. At least one of these mechanisms will be adopted and will be applicable to deployments other than the data center. It is strongly RECOMMENDED that the accepted mechanism be used in conjunction with BGP SPF in data centers. The BGP discovery mechanism should discovery both peer addresses and endpoints for BFD discovery. Additionally, it would be great if there were a heuristic for determining whether the peer is at a tier above or below the discovering BGP speaker (refer to Section 6.2.2).¶
The BGP discovery mechanisms under consideration are [I-D.acee-idr-lldp-peer-discovery], [I-D.xu-idr-neighbor-autodiscovery], and [I-D.ietf-lsvr-l3dl].¶
In order to conserve IPv4 address space and simplify operations, BGP-LS SPF routers in CLOS/Fat-Tree deployments can use IPv6 addresses as peer address. For IPv4 address families, IPv6 peering as specified in [RFC5549] can be deployed to avoid configuring IPv4 addresses on BGP-LS SPF router interfaces. When this is done, dynamic discovery mechanisms, as described in Section 6.5, can used to learn the global or link-local IPv6 peer addresses and IPv4 addresses need not be configured on these interfaces. If IPv6 link-local peering is used, then configuration of IPv6 global addresses is also not required and these IPv6 link-local addresses must then be advertised in the BGP-LS Link Descriptor IPv6 Address TLV (262) [RFC7752].¶
Irrespective of whether or not BGP-LS SPF is used for route calculation, the BGP-LS SPF route advertisements can be used to periodically construct the CLOS/FAT Tree topology. This is especially useful in deployments where an IGP is not used and the base BGP-LS routes [RFC7752] are not available. The resultant topology visibility can then be used for troubleshooting and consistency checking. This would normally be done on a central controller or other management tool which could also be used for fabric data path verification. The precise algorithms and heuristics, as well as, the complete set of management applications is beyond the scope of this document.¶
Since BGP SPF is to be used for the routing underlay and DCI gateway boxes typically have direct or very simple connectivity, BGP external sessions would typically not include the BGP SPF SAFI.¶
The BGP SPF extensions [I-D.ietf-lsvr-bgp-spf] can be used in other topologies and avail the inherent convergence improvements. Additionally, sparse peering techniques may be utilized Section 6.2. However, determining whether or to establish a BGP session is more complex and the heuristic described in Section 6.2.2 cannot be used. In such topologies, other techniques such as those described in [I-D.ietf-lsr-dynamic-flooding] may be employed. One potential deployment would be the underlay for a Service Provider (SP) backbone where usage of a single protocol, i.e., BGP, is desired.¶
In certain scenarios, a BGP node wishes to participate in the BGP SPF topology but never be used for transit traffic. These in include situations where a server wants to make application services available to clients homed at subnets throughout the BGP SPF domain but does not ever want to be used as a router (i.e., carry transit traffic). Another specific instance is where a controller is resident on a server and direct connectivity to the controller is required throughout the entire domain. This can readily be accomplished using the BGP-LS Node NLRI Attribute SPF Status TLV as described in [I-D.ietf-lsvr-bgp-spf].¶
Existing BGP policy including aggregation and prefix filtering may be used in conjunction with the BGP-LS SPF SAFI. When aggregation policy is used, BGP-LS SPF prefix NLRI will be originated for the aggregate prefix and BGP-LS SPF prefix NLRI for components will be filtered. Additionally, link and node NLRI may be filtered and the abstracted by the prefix NLRI.¶
When BGP policy is used with the BGP-LS SPF SAFI, BGP speakers in the BGP-LS SPF routing domain will not all have the same set of NLRI and will compute a different BGP local routing table. Consequently, care must be taken to assure routing is consistent and blackholes or routing loops do not ensue. However, this is no different than if tradition BGP routing using the IPv4 and IPv6 address families were used.¶
No IANA updates are requested by this document.¶
This document introduces no new security considerations above and beyond those already specified in the [RFC4271] and [I-D.ietf-lsvr-bgp-spf].¶
The authors would like to thank Alvaro Retana, Yan Filyurin, and Boris Hassanov for their review and comments.¶