BGP Deterministic Path Forwarding in a CLOS Network
BGP Deterministic Path Forwarding Overview
BGP deterministic path forwarding (DPF) divides a physical fabric into multiple logical fabrics, where different flows are mapped to different logical fabrics to serve the requirements of the flows. Single-hop EBGP's best effort service might not meet all data center flow requirements, especially for drop and latency-sensitive AI-ML flows. BGP DPF selects a path based on the specified logical fabric to ensure optimum link utilization.
BGP DPF is an alternative to RSVP-TE for data center IP fabrics. Data centers usually use single-hop EBGP for IPv4 and IPv6 routing. This is a simple and scalable hop-by-hop EBGP routing that provides a single best effort service for all traffic flows. BGP DPF divides a physical fabric into multiple logical fabrics. This allows both IPv4, and IPv6 traffic to be mapped to different logical fabrics. You can use separate logical fabrics for load balancing across links based on available bandwidth, and with different service level agreements (SLA). You can configure DPF to map flows to multiple logical fabrics to avoid fate sharing.
BGP DPF colors the single-hop EBGP session on each link with a fabric color. For example, if a link belongs to the red fabric, the EBGP session over the link is assigned a red BGP color community. A single-hop EBGP neighbor can be assigned a color at the global, group or neighbor level. A route with no color community assigned can be advertised over any colored or uncolored EBGP sessions.
-
If there is a color mismatch on either side, the receiver marks the route as hidden.
-
All colored routes can be advertised over an uncolored BGP neighbor. To allow certain routes over only certain colors, do not have a mix of uncolored and colored fabrics.
-
You can advertise uncolored routes over all colored fabrics. To advertise matching colored routes over a colored fabric, do not have uncolored routes, except for those carrying light control traffic only.
In Figure 1, Spine1 and Spine 2 are connected to leaf 1 and leaf 2 devices with multiple connections. BGP DPF colors the single-hop EBGP session on each link with a fabric color. Spine 1 belongs to the red fabric, the EBGP sessions over the red link are also colored red. Spine 2 belongs to the blue fabric, the EBGP sessions over the link are also colored blue. BGP advertises IP routes over the EBGP session based on color matching. BGP advertises a route over a red EBGP session if it only has the red color community. BGP advertises a route over a blue EBGP session if it belongs to the blue color community.
Benefits of BGP DPF
-
A lightweight traffic engineering solution for IP fabrics.
-
Correlation of underlay with overlay that isolates large elephant flows from small mice flows.
fabric-color configuration. However, we still support
these features for uncolored routes.