BGP for the Data Center
- BGP Path Hunting and ASN Scheme for Data Centers
- Implementing BGP for the Underlay
- Auto-Discovered BGP Neighbors
- Summary
This chapter is from the book
This chapter is from the book
This chapter is from the book
As described in Chapter 2, “Overview of Data Center Architecture,” modern data centers are built with a scale-out strategy (rather than a scale-up strategy), with predominantly east-west traffic as opposed to the north-south traffic in the traditional three-tier architecture. This shift in strategy was prompted by many factors, including the rise of server virtualization, deployment of high-density server clusters (requiring inter-server communication), new technologies facilitating virtual machine migrations, a shift toward cloud-native applications and workloads, and, more recently, deployment of GPU clusters for artificial intelligence.
In line with this shift in strategy, data center topologies have evolved from a three-tier architecture to a 3-stage Clos architecture (and 5-stage Clos fabrics for large-scale data centers), with the need to eliminate protocols such as Spanning Tree, which made the infrastructure difficult (and more expensive) to operate and maintain due to its inherent nature of blocking redundant paths. Thus, a routing protocol was needed to convert the network natively into Layer 3, with ECMP for traffic forwarding across all available equal cost links. Operational expenditure (OPEX) considerations are equally important as well, since OPEX greatly exceeds capital expenditure (CAPEX) in most IT budgets—the goal should be using a simpler control plane, attempting to reduce control plane interaction as much as possible, and minimizing network downtime due to complex protocols.
In the past, BGP has been used primarily in service provider networks, to provide reachability between autonomous systems globally. BGP was (and still is) the protocol of the Internet, for inter-domain routing. BGP, being a path vector protocol, relies on routing based on policy (with the autonomous system number [ASN] usually acting as a tie-breaker), compared to interior gateway protocols such as Open Shortest Path First (OSPF) and Intermediate System-to-Intermediate System (IS-IS), which use path selection based on a shortest path first logic.
RFC 7938, “Use of BGP for Routing in Large-Scale Data Centers,” provides merit to using BGP with a routed design for modern data centers with a 3-stage or 5-stage Clos architecture. For VXLAN fabrics, external BGP (eBGP) can be used for both the underlay and the overlay. This chapter provides a design and implementation perspective of how BGP is adapted for the data center, specifically with eBGP for the underlay, offering the following features for large-scale deployments:
It enables a simpler implementation, relying on TCP for underlying transport and to establish adjacency between BGP speakers.
Although BGP is assumed to be slower to converge, with minimal design changes and well-known ASN schemes, such problems are nonexistent.
Implementing eBGP for the underlay (for the IPv4 or IPv6 address family) and eBGP for the overlay (for the EVPN address family) using BGP groups in Junos provides a clear, vertical separation of the underlay and the overlay.
Using BGP for both the underlay and overlay provides a simpler operational and maintenance experience. Additionally, eBGP is generally considered easier to deploy and troubleshoot, with internal BGP (iBGP) considered to be more complicated with its need for route reflectors (or confederations) and its best path selection.
Implementing auto-discovery of BGP neighbors using link-local IPv6 addressing and leveraging RFC 8950 (which obsoletes RFC 5549) to transport IPv4 Network Layer Reachability Information (NLRI) over an IPv6 peering for the underlay enables plug-and-play behavior for any new leafs and spines.
BGP Path Hunting and ASN Scheme for Data Centers
Every BGP-speaking system requires an ASN to be assigned to exchange network reachability information with other BGP-speaking systems. An iBGP peering is defined as two BGP speakers with the same ASN peering to each other; an eBGP peering is defined as two BGP speakers with different ASNs peering to each other. For the Internet, publicly owned and assigned ASNs are used (allocated by the Internet Assigned Numbers Authority, or IANA), but this is dangerous for private data centers. One of the most common outages on the Internet is caused by ASN hijacking, in which an organization advertises routes from an ASN that is publicly owned by a different organization or service provider.
For this reason, IANA provides a list of 16-bit and 32-bit private ASNs that organizations can use. The 16-bit private ASNs range from 65412 to 65534, giving only 1023 available ASNs for use. To overcome this limitation, IANA offers 32-bit private ASNs for use as well, providing a much larger range, from 4200000000 to 4294967294. It is imperative that organizations building their own private data centers use ASNs from these private ranges for internal peering.
BGP is designed to route between autonomous systems, where the destination IP prefix is chosen based on the shortest number of AS hops (assuming no policy modification). These AS hops are tracked as part of a BGP attribute called AS_PATH.
In a densely interconnected topology such as a 3-stage Clos network, BGP can suffer from a problem known as path hunting. Path hunting occurs when BGP, on losing a route, hunts for reachability to the destination via all other available paths, not knowing whether the route still exists in the network or not.
Consider the 3-stage Clos network shown in Figure 3-1, with every node assigned a unique ASN from the 16-bit private ASN range.
)
Figure 3-1 Three-stage Clos network with unique ASNs per fabric node
In this topology, leaf1 advertises a subnet x/y to spine1, as shown in Figure 3-2. This route is learned on spine1 with an AS_PATH attribute of [65421]. At the same time, the route is also advertised to spine2, and both spines advertise the route to leaf2 and leaf3.
BGP, by default, only advertises the best route to its neighbors. When leaf2 and leaf3 receive this route from both spine1 and spine2, they must elect one path as the best path. With no policy modification, the best path is chosen based on the shortest AS_PATH attribute, but in this case, the AS_PATH length is the same because the route received from spine1 will have an AS_PATH of [65500 65421] and the route received from spine2 will have an AS_PATH of [65501 65421]. Eventually, this tie-breaker is broken by selecting the oldest path. Assuming the elected best path is via spine2 (since it is the oldest path), leaf2 and leaf3 advertise this route to their eBGP peer list, which, in this case, consists only of spine1 (the route cannot be advertised back to spine2 because it originally sent the route that was elected as the best route).
)
Figure 3-2 Subnet x/y advertised to spine1 and spine2 by leaf1
Thus, spine1 receives this route back from leaf2 and leaf3. At this point, spine1 has multiple paths available to reach subnet x/y advertised by leaf1; however, only the direct path (via leaf1) is selected as the best path, since it has the shortest AS_PATH length (again, assuming there are no policy modifications), as shown in Figure 3-3.
)
Figure 3-3 Routing table on spine1 showing all available paths for subnet x/y
When spine1 loses its best path to subnet x/y, which is via leaf1 (leaf1 goes down or withdraws the route), it hunts for an alternate best path from all available paths. At the same time, spine1 also sends a BGP withdraw to its neighbors, informing them of the lost route via leaf1 for subnet x/y. Eventually, once all withdraws have converged and the subnet has been fully purged from the network, spine1 has no available paths for it, and the route is removed from its routing table.
While this path-hunting behavior might appear to be a minor problem, it becomes increasingly problematic as the fabric size increases with more leafs, creating many alternate paths to hunt through. Thus, to avoid this problem, and to speed up BGP convergence, either of the following two methodologies can be followed, with the same end goal of ensuring that the spines do not learn alternate, suboptimal routes reflected from other leafs:
Use an ASN scheme, leveraging eBGP’s built-in loop-prevention mechanism of dropping updates that include its own ASN in its AS_PATH list. This is the default BGP behavior, and you do need to configure any additional policies for this.
Use routing policies to prevent spines from accepting routes that were originally advertised by any other spine.
This ASN scheme is represented in Figure 3-4.
)
Figure 3-4 BGP ASN scheme for a 3-stage Clos fabric to avoid path hunting with same ASN on all spines
For a 5-stage Clos fabric, the ASN scheme mandates that all spines within a pod share the same ASN, but spines across pods have unique ASNs. Additionally, all leafs in each pod are assigned a unique ASN, while all superspines share the same ASN. This ASN scheme is represented in Figure 3-5.
Thus, for a 3-stage or 5-stage Clos fabric, with the ASN schemes shown in Figures 3-4 and 3-5, BGP path hunting is natively prevented.
The second methodology uses an ASN scheme in which all fabric nodes use a unique ASN, and routing policies are used to control how routes are advertised back to the spines to prevent BGP path hunting. In this case, as the spines advertise routes to the leafs, they are tagged with a BGP community using an export policy. On the leafs, an export policy is used to prevent the advertisement of routes with this BGP community from being sent back to the spines, thus preventing the existence of route state on the spines that can lead to path hunting. This is shown in Figure 3-6.
)
Figure 3-5 BGP ASN scheme for a 5-stage fabric to avoid path hunting
)
Figure 3-6 Routing policy logic to prevent path hunting
This implementation, while more complex and requiring additional operational overhead in the form of policy configuration, is necessary in certain designs where external devices are connected to the fabric for inter-VRF routing. Consider the topology shown in Figure 3-7, where the same ASN is used for both spines and a firewall is connected to leaf3 for inter-VRF routing.
)
Figure 3-7 Firewall connected to fabric leaf for inter-VRF routing
In Figure 3-7, leaf1 is configured with an IP VRF v10, which includes an IPv4 subnet 172.16.10.0/24, and leaf2 is configured with an IP VRF v20, which includes an IPv4 subnet 172.16.20.0/24. The firewall has a BGP peering to leaf3 over both these IP VRFs to leak routes from one VRF to another.
The IPv4 subnet 172.16.10.0/24 is advertised by leaf1 toward leaf3, and eventually to the firewall, with an AS_PATH list of [65423 65500 65421], as shown in Figure 3-8.
)
Figure 3-8 AS_PATH attribute as a prefix, originated by leaf1, is advertised toward firewall
The firewall “leaks” this route into IP VRF v20 by advertising it to the VRF-specific BGP neighbor on leaf3. Thus, leaf3 receives this in IP VRF v20 and advertises it to the rest of the fabric via the spines. However, when the spines receive this BGP update, they drop it because their local ASN is present in the AS_PATH list and BGP loop prevention rules indicate that such an update must be dropped. This is shown in Example 3-1, with BGP debugs on spine1.
Example 3-1 Spines dropping BGP update due to AS loop prevention rules
Jan 14 17:34:26.497233 BGP RECV 192.0.2.13+179 -> 192.0.2.101+61507 Jan 14 17:34:26.497273 BGP RECV message type 2 (Update) length 128 Jan 14 17:34:26.497369 BGP RECV Update PDU length 128 Jan 14 17:34:26.497452 BGP RECV flags 0x40 code Origin(1): IGP Jan 14 17:34:26.497517 BGP RECV flags 0x40 code ASPath(2) length 22: 65423 65510 65423 65500 65421 Jan 14 17:34:26.497550 BGP RECV flags 0xc0 code Extended Communities(16): 2:502:502 encapsulation:vxlan(0x8) router-mac:2c:6b:f5:75:70:f0 Jan 14 17:34:26.497561 BGP RECV flags 0x90 code MP_reach(14): AFI/SAFI 25/70 Jan 14 17:34:26.497577 BGP RECV nhop 192.0.2.13 len 4 Jan 14 17:34:26.497650 BGP RECV 5:192.0.2.14:502::0::172.16.10.0::24/248 (label field value 0x2906 [label 656, VNID 10502]) (esi 00:00:00:00:00:00:00:00:00:00) Jan 14 17:34:26.497661 End-of-Attributes Jan 14 17:34:26.497910 As loop detected. Rejecting update *snip*
Figure 3-9 shows a visual representation of the same behavior.
)
Figure 3-9 BGP update dropped on spine1 due to local ASN 65500 in AS_PATH
These problems can be circumvented by allowing the same ASN to be present in the AS_PATH attribute using several configuration options in Junos or by using an ASN scheme where each spine is assigned a unique ASN. Intent-based networking systems such as Juniper Apstra take away the complexity of implementing such an ASN scheme by automating and orchestrating the configuration of necessary policies to prevent path hunting (since that is the prevailing problem when each spine is assigned a unique ASN), with no requirement of operator intervention, while also facilitating designs as shown in Figure 3-7.