Quick reroute in Clos networks
Techniques for implementing quick reroute in a Clos network comprising a plurality of leaf network devices and a plurality of spine network devices are provided. In one set of embodiments, a first spine network device can configure an Equal-Cost Multi-Path (ECMP) group that comprises the plurality of leaf network devices except for a first leaf network device. The first spine network device can further detect a connectivity failure between the first spine network device and the first leaf network device. In response to the failure, the first spine network device can update one or more routes in a routing table, where the one or more routes are associated with one or more hosts that are connected to the first leaf network device, and where updating each route in the one or more routes comprises changing the route to point to the ECMP group as a next hop.
1 . A method for implementing quick reroute in a Clos network comprising a leaf layer with a plurality of leaf network devices and a spine layer with a plurality of spine network devices, the method comprising:
configuring, by a first spine network device in the plurality of spine network devices, a first set of Equal-Cost Multi-Path (ECMP) groups on the first spine network device, the first set of ECMP groups including a first ECMP group that comprises the plurality of leaf network devices except for a first leaf network device;
detecting, by a first spine network device, a connectivity failure between the first spine network device and the first leaf network device; and
in response to the detecting, updating, by the first spine network device, one or more routes in a routing table of the first spine network device, wherein the one or more routes are associated with one or more hosts that are connected to the first leaf network device, and wherein updating each route in the one or more routes comprises changing the route to point to the first ECMP group as a next hop.
2 . The method of claim 1 wherein each ECMP group in the first set of ECMP groups comprises the plurality of leaf network devices except for a particular leaf network device in the plurality of leaf network devices.
3 . The method of claim 1 wherein the particular leaf network device is a different leaf network device for said each ECMP group.
4 . The method of claim 1 wherein the updating causes a network packet that is received by the first spine network device and is destined for a host in the one or more hosts to be rerouted to a second leaf network device in the plurality of leaf network devices.
5 . The method of claim 4 wherein upon receiving the network packet, the second leaf network device reroutes the network packet to a second spine network device in the plurality of spine network devices that is connected to the first leaf network device.
6 . The method of claim 5 wherein upon receiving the network packet, the second spine network device forwards the packet to the first leaf network device.
7 . The method of claim 4 wherein the second leaf network device has configured thereon:
a second set of ECMP groups, each ECMP group in the second set of ECMP groups including the plurality of spine network devices except for a particular spine network device; and
a policy-based routing (PBR) policy including a first group of higher priority entries and a second group of lower priority entries.
8 . The method of claim 7 wherein each entry in the first group of higher priority entries is configured to match packets whose next hop is a host connected to the second leaf network device and to refrain from overriding the next hop of the matched packets.
9 . The method of claim 7 wherein each entry in the second group of lower priority entries is configured to match packets that are received on a source port connected to a source spine network device and to override a next hop of the matched packets to point to an ECMP group in the second set of ECMP groups that excludes the source spine network device.
10 . The method of claim 9 wherein said each entry in the second group of lower priority entries is further configured to copy the matched packets to a central processing unit (CPU) of the second leaf network device.
11 . The method of claim 7 wherein the second leaf network device further has configured thereon a third set of ECMP groups, each ECMP group in the third set of ECMP groups including all spine network devices in the plurality of spine network devices that the second leaf network device can use to directly reach a particular leaf network device, except for a particular spine network device in said all spine network devices.
12 . The method of claim 11 wherein the PBR policy further includes a third group of entries between the first group of higher priority entries and the second group of lower priority entries, each entry in the third group of entries being configured to match packets that are received on a source port connected to a source spine network device and to override a next hop of the matched packets to point to an ECMP group in the third set of ECMP groups that excludes the source spine network device.
13 . The method of claim 12 wherein said each entry in the third group of entries is further configured to copy the matched packets to a CPU of the second leaf network device.
14 . The method of claim 10 wherein, upon determining from the matched packets that the first leaf network device has lost connectivity to every spine network device in the plurality of spine network devices, the CPU of the second leaf network device programs a data plane of the second leaf network device to drop all packets destined for hosts connected to the first leaf network device.
15 . The method of claim 1 wherein the Clos network further comprises a super-spine layer with a plurality of super-spine network devices that are connected to the plurality of spine network devices.
16 . A spine network device in a Clos network comprising a plurality of spine network devices and a plurality of leaf network devices, the spine network device comprising:
one or more processors; and
a computer-readable storage medium having stored thereon program code that, when executed by the one or more processors, causes the one or more processors to:
configure a set of Equal-Cost Multi-Path (ECMP) groups on the spine network device, the set of ECMP groups including an ECMP group that comprises the plurality of leaf network devices except for a first leaf network device;
detect a connectivity failure between the spine network device and the first leaf network device; and
in response to the detecting, update one or more routes in a routing table of the spine network device, wherein the one or more routes are associated with one or more hosts that are connected to the first leaf network device, and wherein updating each route in the one or more routes comprises changing the route to point to the ECMP group as a next hop.
17 . A method for implementing quick reroute in a Clos network comprising a leaf layer with a plurality of leaf network devices, a spine layer with a plurality of spine network devices, and a super-spine layer with a plurality of super-spine network devices, the method comprising:
configuring, by a first spine network device in the plurality of spine network devices, a first set of Equal-Cost Multi-Path (ECMP) groups on the first spine network device, the first set of ECMP groups including a first ECMP group that comprises the plurality of super-spine network devices;
detecting, by a first spine network device, a connectivity failure between the first spine network device and a first leaf network device in the plurality of leaf network devices; and
in response to the detecting, updating, by the first spine network device, one or more routes in a routing table of the first spine network device, wherein the one or more routes are associated with one or more hosts that are connected to the first leaf network device, and wherein updating each route in the one or more routes comprises changing the route to point to the first ECMP group as a next hop.
18 . The method of claim 17 wherein the updating causes a network packet that is received by the first spine network device and is destined for a host in the one or more hosts to be rerouted to a super-spine network device in the plurality of super-spine network devices.
19 . The method of claim 18 wherein upon receiving the network packet, the super-spine network device reroutes the network packet to a second spine network device in the plurality of spine network devices that is connected to the first leaf network device.
20 . The method of claim 19 wherein upon receiving the network packet, the second spine network device forwards the packet to the first leaf network device.