Thursday, January 22, 2015

CCIE OCG: EIGRP (Chapter 8), Part IV

EIGRP Named Mode

  • Starting with IOS Release 15.0(1)M.
  • Earlier implementation configured EIGRP processes directly by their AS numbers separately for IPv4 and IPv6, also called the classic mode or the autonomous system mode.
  • The named mode is preferred.
  • The classic mode remains for backwards compatibility.
  • The reason for the update was to produce a better unified and more consistent configuration interface.
  • A single EIGRP process configuration consists of one or more address family sections, each of them specifying details of EIGRP operation for the particular address family.
  • Even former per-interface commands are within the context of the named mode.
  • The entire EIGRP configuration is located in a single place.
  • If an EIGRP instance is configured in named mode, all EIGRP commands outside the named mode will be ignored.
  • The classic mode and named mode are just two different ways of how EIGRP is configured. They do not constitute two different versions of EIGRP.
  • Three building blocks:
  1. Address Family (AF) section: Mandatory section that specifies the particular address family for which an EIGRP instance shall be started. The autonomous system number is part of the AF section.
  2. Per-AF-interface section: Optional section holds EIGRP settings pertaining to the specified interface and AF. One per-AF-interface section can be created for each routed interface or subinterface. In case both af-interface default and a specific af-interface define the same setting, the interface-specific section is preferred.
  3. Per-AF-topology section: A section presented inside a particular AF, related to the support of Multi Topology Routing (MTR) in EIGRP. The topology base will always be present in the configuration, even if the IOS has no support for multiple routing topologies.

router eigrp CCIE
 address-family ipv4 unicast autonomous-system 1
   af-interface default
     hello-interval 1
     hold-time 3
   exit-af-interface
   af-inteface Loopback0
     passive-interface
   exit-af-interface
   topology base
     maximum-paths 6
     variance 4
   exit-af-topology
   network 10.0.0.1 0.0.0.0
   network 10.255.255.1 0.0.0.0
 exit-address-family
!
 address-family ipv6 unicast autonomous-system 1
   af-interface default
     shutdown
     exit-af-interface
   af-interface Loopback0
     no shutdown
     exit-af-interface
   af-interface FastEthernet0/0
     no shutdown
     exit-af-interface
   topology base
     timers active-time 1
     exit-af-topology
   exit-address-family


  • As soon as the IPv6 EIGRP address family is configured, it immediately runs on all IPv6-enabled interfaces.
  • The named mode is entered using the router eigrp <name> command, where <name> is an arbitrary text name (a virtual instance name) of the EIGRP process.
  • The process name is not sent in EIGRP messages; it is locally significant value and is never compared to process names on other routers.
  • Each named EIGRP process can hold only a single instance for an address family.
  • If it is necessary to run two IPv4 EIGRP instances (for example, AS number 1 and AS number 64512), each of them must be placed into a separate EIGRP process with a unique name.
  • Also, two or more distinct named EIGRP processes cannot run the same address family instance with the same AS number.
  • One-to-one correspondence between an EIGRP named process and an address family instance with a particular AS number.
  • The AS numbers for individual address families in the same EIGRP process do not need to match.
  • In the classic mode, an IPv6 EIGRP process was shut down by default after configuring it, and no shutdown was necessary to actually start it.
  • In the named mode, the opposite is true. As soon as IPv6 address family is configured, it automatically adopts all interfaces on which IPv6 has been enabled (even link-local addresses).
  • Typically, use the af-interface default to first keep the IPv6 address family instance off all interfaces, and only activate it on selected interfaces later.
  • Notice the difference between passive-interface default and af-interface default shutdown.
  • passive-interface: no EIGRP adjacencies will be formed, but its global prefixes will still be advertised over other interfaces.
  • shutdown: completely igonored by EIGRP, not forming any adjacencies over it and also not advertising any of its prefixes.

Address Family Section

  • Configurations specific to the EIGRP process itself.
  • Commonly used commands include network and neighbor statements.

Address Family configuration commands:
  • af-interface: enter Address Family interface configuration.
  • default: set a command to its defaults.
  • eigrp: EIGRP Address Family specific commands.
  • exit-address-family:exit Address Family configuration mode.
  • help: description of the interactive help system.
  • maximum-prefix: maximum number of prefixes acceptable in aggregate.
  • metric: modify metrics and parameters for advertisement.
  • neighbor: specify an IPv4 neighbor router.
  • network: enable routing on an IP network.
  • shutdown: shutdown address family.
  • timers: adjust peering based timers.
  • topology: topology configuration mode.

Per-AF-interface Configuration Section

  • EIGRP interface-specific commands are applied, except bandwidth and delay.
  • No summary-address command under the default interface.

Address Family Interfaces configuration commands:
  • add-paths: advertise add paths.
  • authentication: authentication subcommands.
  • bandwidth-percent: set percentage of bandwidth percentage limit.
  • bfd: enable Bidirectional Forwarding Detection.
  • dampening-change: percent interface metric must change to cause update.
  • dampening-interval: time in seconds to check interface metrics.
  • default: set a command to its defaults.
  • exit-af-interface: exit from Address Family Interface configuration mode.
  • hello-interval: configure hello interval.
  • hold-time: configure hold time.
  • next-hop-self: configure EIGRP next-hop-self.
  • no: negate a command or set its defaults.
  • passive-interface: suppress address updates on an interface.
  • shutdown: disable Address-Family on interface.
  • split-horizon: perform split horizon.
  • summary-address: perform address summarization.

Per-AF-Topology Section

  • Relates to the support of multiple routing topologies in EIGRP.
  • Outside the scope of CCIE Routing and Switching certification.
  • A topology is defined as a subset of routers and links in a network for which a separate set of routes is calculated.
  • The base topology is the default routing environment that exists prior to enabling MTR.
  • Any additional topologies are known as class-specific topologies and are a subset of the base topology.
  • Each class-specific topology carries a class of traffic and is characterized by an independent set of Network Layer Reachability Information (NLRI) that is used to maintain separate routing tables and FIB databases.
  • The router can perform independent route calculation and forwarding for each topology.
  • Multiple topologies can be used to segregate traffic classes (for example, route data, voice, and video over different links).
  • MTR is not equivalent to VRF, because they share the common address space, and they are not intended to provide address conservation or reuse.
  • On routers without MTR support, only the topology base will be available.

Address Family Topology configuration commands:
  • auto-summary: enable automatic network number summarization.
  • default: set a command to its defaults.
  • default-information: control distribution of default information.
  • default-metric: set metric of redistributed routes.
  • distance: define an administrative distance.
  • distribute-list: filter entries in EIGRP updates.
  • eigrp: EIGRP specific commands.
  • exit-af-topology: exit from Address Family Topology configuration mode.
  • maximum-paths: forward packets over multiple paths.
  • metric: modify metrics and parameters for advertisement.
  • no: negate a command or set its defaults.
  • offset-list: add or subtract offset from EIGRP metrics.
  • redistribute: redistribute IPv4 routes from another routing protocol.
  • snmp: modify SNMP parameters.
  • summary-metric: specify summary to apply metric/filtering.
  • timers: adjust topology specific timers.
  • traffic-share: how to compute traffic share over alternate paths.
  • variance: control load balancing variance.

  • show commands have also been updated.
  • Instead of show ip eigrp ..., the new syntax show ip eigrp address-family ipv4 is used.
  • The older show commands are still acceptable for named mode.


Monday, January 19, 2015

CCIE OCG: EIGRP (Chapter 8), Part III

Diffusing Update Algorithm (DUAL)

  • A convergence algorithm that replaces the Bellman-Fordman algorithm used by other distance-vector protocols.
  • DUAL uses a concept of diffusing computations to perform distributed shortest-path computation while maintaining freedom from loops during those calculations.
  • DUAL is at the center of the EIGRP routing protocol.

Topology Table

  • The central data store of an EIGRP process.
  • Unconvetional naming because EIGRP has no information about the network's topology per se.
  • EIGRP stores its entire routing information in the topology table, including:
  1. The prefix of each known destination network (address/netmask).
  2. Feasible Distance of the destination network
  3. Address of each neighboring router that advertised the destination network, including the egress interface toward the neighbor.
  4. Metrics of the destination network as advertised by each neighbor, and the resulting metrics of the path through that neighbor
  5. State of the destination network
  6. Additional information about the network (various internal flags, network type and origin, and others)
  • Populated and updated by locally injected networks (directly connected EIGRP interfaces, routes redistributed locally) and by contents of received EIGRP messages.
  • The least total cost and loop-free paths are installed into the routing table.
  • Each network in the topology table is associated with a state.
  • Passive meaning that the shortest path to the network has already been found, and EIGRP is satisfied with it.
  • Active meaning that EIGRP is currently actively involved in a search for a new shortest path.
  • In a stable topology, all routes shall be in the Passive state.
  • The Active state is always related to the router sending Query packets, asking its neighbors for cooperation in the search for a new path.
  • While in an Active state, the router is prohibited from modifying the routing table entry (not removed or its next-hop changed).
  • Whenever a router needs to select a new shortest path and the neighbor providing that path can be proven not to create a routing loop, the route stays in the Passive state because the router already has all the information to make a correct choice.
  • If the neighbor providing the least-cost path cannot be guaranteed not to create a routing loop, or if no such neighbor exists, the route will need to enter the Active state.

Computed, Reported, and Feasible Distances, and Feasibility Condition

R1# show ipv6 eigrp topology all-links
IPv6-EIGRP Topology Table for AS(1)/ID(10.255.255.1)
Codes: P - Passive, A - Active, U - Update, Q - Query, R - Reply,
r - reply Status, s - sia Status
P 2001:DB8:CC1E::/64, 1 successors, FD is 2560, serno 4
via FE80::2 (2560/256), Serial1/0
via FE80::3 (5120/1280), Serial1/1
via FE80::4 (4096/3072), Serial1/2


  • The "via" lines describe a possible route to the destination through a particular neighbor.
  • Contains the next-hop's IPv6 address, egress interface, and two numbers enclosed in parentheses.
  • The number after the slash sign is called Reported Distance (RD).
  • RD is the neighbor's distance to the destination as reported in an EIGRP packet received from that neighbor.
  • The number before the slash sign is called the Computed Distance (CD).
  • CD shows the total metric of reaching the destination over the particular neighbor.
  • The CD is computed as the RD of neighbor plus the cost of the link between R1 and the neighbor.
  • The Feasible Distance (FD) is a record of the lowest known distance since the last transition from the Active to Passive state.
  • In other words, FD is a historical record, or a historical copy, of the smallest known CD toward a particular destination since the last Active-to-Passive transition.
  • FD is an internal value that is used by EIGRP to select loop-free paths, but its value is never advertised in any EIGRP packets.
  • FD describes the metric of the best path to the destination this router has known (the smallest known distance to a destination since the last time the route went Passive).
  • Any neighbor whose Reported Distance is strictly smaller than this router's Feasible Distance cannot form a routing loop.
  • The Feasibility Condition states that every neighbor satisfying the inequality RD < FD provides a loop-free path, also sometimes called the Source Node Condition.
  • The FC is a sufficient condition, not a necessary condition.
  • Every neighbor satisfying the FC provides a loop-free path, but not every loop-free path satisfies the FC.
  • The FC basically splits all neighbors into two groups: 1) neighbors that are guaranteed to provide a loop-free path, and 2) all other neighbors about which the router cannot be sure.
  • Feasible Successor: all neighbors that pass the FC and thus are safe to use as next-hops.
  • Successor: the Feasible Successor with the least CD to the destination.
  • show ip eigrp topology displays Successors and Feasible Successor.
  • show ip eigrp topology all-links displays all EIGRP routes.

 

Local and Diffusing Computations in EIGRP

  • A topology change occurs whenever the distance to a network changes or a new neighbor comes online that advertises the network.
  • The distance change can be detected either through receiving an EIGRP packet, or because a local interface metric has changed.
  • Also, a neighbor failure is processed by setting the CD/RD of all networks reachable through that neighbor to infinity.

  • Convergence after a topology change:
  1. The Feasible Successor providing the least CD is made the new Successor.
  2. If the CD over the new Successor is less than the current FD, the FD will be updated to the new CD; otherwise it stays as its current value.
  3. The routing table is updated to point toward the new Successor.
  4. If the current distance to the destination has changed as a result of the new Successor, an Update packet is sent to all neighbors, advertising the router's updated distance.
  • This is a local computation, performed solely by using information already stored in the topology table, and not involving neighbors. The route has remained in the Passive state.

  • If, after detecting a topology change, the router finds out that the new shortest path is provided by a neighbor that is not a Feasible Successor, the router cannot use this neighbor right away because it could cause a routing loop.
  • The router performs a diffusing computation:
  1. The entry in the routing table is locked. It must not be removed nor its next-hop changed until the diffusing computation is finished and the route has been moved to the Passive state again.
  2. The FD is set to the current (possibly increased) CD through the current unchanged Succesor. If the router ever needs to advertise its distance to the network while in the Active state, it will use the current CD.
  3. The network is put into the Active state and the router sends out a Query packet to all its neighbors. This Query packet contains the Active network's prefix and the router's current CD toward it.

  • Each neighbor receiving a Query packet will process it by updating its own topology table using the distance information advertised in the Query and reevaluating its own choice of Successors and Feasible Successors.
  • Two possibilities: 1) the neighbor still has its own Successor or Feasible Success, or 2) the information in the Query causes the neighbor to stop considering the path through its current Successor the shortest path available and it has no Feasible Successors.
  • 1) When the neighbor still has a Successor, it will simply send back a Reply packet, indicating the neighbor's current distance to the destination (performing its local computation if necessary). The neighbor did not become engaged in the diffusing computation, and did not need to put the network in the Active state itself.
  • 2) The neighbor will itself join the diffusing computation, send out its own Query packet, and advertise its own current distance through its current Successor. As a result, the wave of Query messages propagates through the part of the network that is affected by the change. (EIGRP marketing term "partial, bounded updates": partial = only the changed information, bounded = only in the affected part of the network).

  • In the Active state, a router must wait for a Reply packet from each Queried neighbors.
  • Until then, the route remains in the Active state, and its routing table is unchanged.
  • Only after all Reply packets are received, the router can put the route back in the Passive state, simply choose the shortest path neighbor skipping the FC check, and reinitialize the FD to the CD of the selected neighbor.
  • The crucial information is simply the sender's current distance to the destination, informing its receiver about the packet originator's distance to the destination.
  • Whether the packet causes its receiver to go Active depends exclusively on how the information in the message impacts the receiver's choice of the shortest path and FC check.

  • Whenever EIGRP detects a topology change, it first records the change into the topology table and updates the RD and CD of the neighbor that advertised the change (received EIGRP message) or was influenced by it (link metric change).
  • From among all neighbors that advertise the network, EIGRP identifies the on that provides the least CD, taking into account the updated CDs. Note that the FC is not invoked at this point.
  • Only after identifying the neighbor offering the least CD, EIGRP verifies whether this neighbor meets the FC and is a Feasible Successor. If it is, EIGRP will promote it to the Successor and start using it right away. If the neighbor does not meet the FC, EIGRP will put the route into the Active state and send out Queries.

Before convergence.
R1# show ipv6 eigrp topology all-links
! Legend removed for brevity
P 2001:DB8:CC1E::/64, 1 successors, FD is 2048, serno 9
via FE80::2 (2560/256), Serial1/0
via FE80::3 (5120/1280), Serial1/1
via FE80::4 (4096/3072), Serial1/2



After convergence.
R1# show ipv6 eigrp topology
! Legend removed for brevity
P 2001:DB8:CC1E::/64, 1 successors, FD is 4096
via FE80::4 (4096/3072), Serial1/2
via FE80::3 (5120/1280), Serial1/1


R1# show ipv6 route eigrp
! Legend removed for brevity
D 2001:DB8:CC1E::/64 [90/4096]
via FE80::4, Serial1/2



DUAL FSM

  • Assuming a router to efficiently compute a new path to a destination with no other topological changes taking place is a strong assumption.
  • EIGRP uses a control mechanism called the Diffusing Update Algorithm, or DUAL, to take care of handling multiple topology changes occurring during a single diffusing computation.

  • A0: Local Origin with Distance Increase
  • A1: Local Origin
  • A2: Multiple Origins
  • A3: Successor Origin
  • CD = Computed Distance
  • FC = Feasibility Condition

Summarization of DUAL FSM:
  • Unless a change in distance occurs such that the least CD neighbor fails to meet the FC, the route remains Passive.
  • If a Query is received from the current Successor and the least-cost CD neighbor fails to meet the FC, the route will enter the A3 active state (Successor Origin Active state). The router will send out Queries and wait for Replies. The last Reply allows the router to transition back to the Passive state, reinitialize the FD, and choose the least-cost CD neighbor as the new Successor.
  • If a distance change is not caused by a Query from a Successor (receiving an Update, changing an interface metric, or losing a neighbor) and the least CD neighbor does not meet the FC, the route will enter A1 active state (Local Origin Active state). The router will send out Queries and wait for Replies. The last reply allows the router to transition back to the Passive state, reinitialize the FD, and choose the least-cost CD neighbor as the new Successor.
  • If during the stay in the A 3 (Successor Origin) or A 1 (Local Origin) active states, another distance increase caused by other means than the Successor’s Query is detected, another topology change during the diffusing computation has occurred. Because the router cannot advertise this updated distance while it is in the Active state, other routers might not be informed about it and their Replies might not take this new increased distance into account. The state is changed from A3 (Successor Origin) to A2 (called Multiple Origins ), or from A1 (Local Origin) to A0 (no official name; we will call it Local Origin with Distance Increase ) states. In A2 or A0 states, the router waits to receive all remaining Replies. When the last Reply arrives, the router will first check whether the neighbor providing the least Computed Distance passes the Feasible Condition check using the Feasibility Distance value set when the route entered the Active state. If it does, the route becomes Passive again, and the neighbor is chosenas the Successor. If it does not, however, the route will return from A0 (Local Origin with Distance Increase) to A1 (Local Origin) or from A2 (Multiple Origins) to A3(Successor Origin) and the router will commence another diffusing computation by again sending a Query. 
  • If during the stay in A 1 (Local Origin) or A 0 (Local Origin with Distance Increase) active states a Query from the Successor is received, another topology change during the diffusing computation has occurred. Because the router cannot advertise this updated distance while it is in the Active state, other routers might not be informed about it and their Replies might not take this new increased distance into account. The state is changed to A2 (Multiple Origins) and then proceeding from that state just like in the previous case.

Stuck-In-Active State

  • In the Active state, EIGRP must wait for all its neighbors to send back a Reply before it can conclude the diffusing computation itself, make a new best-path selection, and start sending its own Replies.
  • A route in the Active is dependent on the entire chained sequence of routers that have become Active as a result of its Query.
  • There are several reasons why a neighbor might not respond to a Query:
  1. The CPU is overloaded and the router cannot respond or even process the Query.
  2. Quality issues on the link are causing packets to be lost.
  3. Low-bandwidth links are congested and packets are being delayed or dropped.
  4. The network is excessively large or complex, either requiring the Query to propagate to a significant depth or causing an inordinate number of prefixes to be impacted by a single link or node failure.
  • Active timer: starts when the first Query is sent out. By default, 3 minutes (1-65535 minutes). Set to infinity using timers active-timer.
  • If all expected Replies are not received before the Active timer expires, the route in question will be designated as Stuck-In-Active (SIA).
  • The neighbor(s) that did not Reply will be removed from the neighbor table and their adjacencies torn down.
  • Instability ensues. All networks learned from that neighbor will be flushed and possibly learned again after the neighbor comes back up within the Hello interval time.
  • Difficult to located the problem origin.
  • If the neighbor does not respond to a Query message with its Reply within half of the Active timer time, the router will send the neighbor a SIA-Query message.
  • The SIA-Query: "Are you still working on my Query?"
  • If the neighbor is able to receive and process this SIA-Query, it will immediately respond with the SIA-Reply message.
  • SIA-Reply: "Yes, I'm still waiting for my own neigbors to send me Replies." or "No, the computation is finished. This is my current metric to the destination."
  • Receiving an SIA-Reply allows the Active timer to be reset, giving the diffusing computation additional time to complete.
  • At most three SIA-Queries can be sent, each after half of the Active timer.
  • If the diffusing computation is not finished by the time the third SIA-Query was replied to by an SIA-Reply and half of the Active timer expired again, the adjacency to the neighbor will be dropped.
  • The same will happen if an SIA-Query is not responded to by an SIA-Reply within the next half of the Active timer.
  • With the default settings, the maximum time for SIA = 4 x 90 = 360 seconds (90 seconds for the first SIA-Query, plus each SIA-Query buying another 90 seconds).

  • It is possible for a router to receive a Query for a destination while it is Active for that destination.
  • Some sources claim that this situation causes the SIA states. In reality, this is a gross misunderstanding.
  • When a router enters the Active state, it sends out a Query indicating its current distance to the destination after the topology change that triggered the transition to the Active state.
  • If, during the Active state, the router receives another Query for this destination, it simply sends back a Reply packet immediately, claiming exactly the same distance as originally advertised in its own Query packet.
  • Any deadlock scenario is thereby averted.
  • Proper hierarchical network design coupled with judicious use of passive interfaces, appropritate route filtering and/or summarization and the EIGRP Stub feature are the key tools that help limit the probability of an SIA state occurrence to an absolute minimum.

Thursday, January 15, 2015

CCIE OCG: EIGRP (Chapter 8), Part II

EIGRP Packet Format

  • Carried directly in IP packets, using protcol number 88.
  • The maximum length of an EIGRP packet is derived from the maximum IP MTU on the particular interface - typically 1500 bytes for the entire IP packet, leaving 1480 bytes for the EIGRP packet itself.
  • Each EIGRP packet carries a 20-byte header, followed by a variably sized body (TLVs).
  • There is no clearly delineated RTP header.
  • Instead, the Flags, Sequence Number, and Acknowledgement number fields provide most of the RTP functionality (also specific TLVs).

Version: 4-bit field used to indicate the protocol version of the originating EIGRP process. Has not changed since its release and is set to 2.
Opcode: 4-bit field that specifies the EIGRP packet type. 1 = Update, 3 = Query, 4 = Reply, 5 = Hello/Ack, 10 = SIA-Query, 11 = SIA-Reply.
Checksum: 24-bit field used to run a sanity check on the EIGRP packet, excluding the IP header.
Flags: 32-bit field indicating specific flags. 0x1 = Init (used during initial adjacency buildup), 0x2 = Conditional Receive (used by RTP to allow this message to be received only by a subset of receivers), 0x4 = Restart (indicates that a router has restarted), 0x8 = End-of-Table (indicates that the transmission of the entire EIGRP database is complete).
Sequence: 32-bit field that contains a sequence number used by RTP for orderly delivery of reliable packets.
Acknowledgement: 32-bit field used by RTP that contains the sequence number of the last packet heard from the neighbor. A Hello packet with a non-zero ACK field is an ACK rather than a Hello. Acknowledgements are never multicasted.
Virtual Router ID: 16-bit field identifying the virtual router this packet is associated with. 0x1 = Unicast Address Family, 0x2 = Multicast Address Family, 0x8000 = Unicast Service Address Family.
Autonomous System Number: 16-bit field that identifies the number of the EIGRP domain.
Type-Length-Value: Field used to carry route entries as well as provide EIGRP DUAL information. Several different types:
  • 0x0001 EIGRP Parameters ( General TLV Types )
  • 0x0002 Authentication Type ( General TLV Types )
  • 0x0003 Sequence ( General TLV Types )
  • 0x0004 Software Version ( General TLV Types )
  • 0x0005 Next Multicast Sequence ( General TLV Types ) 
  • 0x0102 IPv4 Internal Routes ( IP-Specific TLV Types )
  • 0x0103 IPv4 External Routes ( IP-Specific TLV Types )
  • 0x0402 IPv6 Internal Routes ( IP-Specific TLV Types )
  • 0x0403 IPv6 External Routes ( IP-Specific TLV Types )
  • 0x0602 Multi Protocol Internal Routes ( AFI-Specific TLV Types )
  • 0x0603 Multi Protocol External Routes ( AFI-Specific TLV Types )

EIGRP Packets

  • EIGRP uses seven different packet types:
  1. Hello
  2. Acknowledgement
  3. Update (RTP) 
  4. Query (RTP)
  5. Reply (RTP)
  6. SIA-Query (RTP)
  7. SIA-Reply (RTP)
  • The RTP marked packets are also called reliable packets because EIGRP makes sure that they are all reliably delivered and in proper order.
  • Statistics about all sent and received EIGRP packets can be obtained using the show ip eigrp traffic command.

Hello

  • Periodic Hellos sent once EIGRP is enabled on an interface.
  • Used to identify neighbors, verify compatible configuration (common IP subnet, the same AS number, K values, and optional authentication), and serve as a keepalive mechanism between neighbors.
  • Sent to 224.0.0.10 in IPv4, and FF02::A in IPv6.
  • If static neighbors are configured, Hello packets are sent as unicasts to the neighbors explicitly configured address.
  • The default Hello interval is 5 seconds.
  • On NBMA with 1544 kbps or less, the default Hello interval is 60 seconds.
  • Opcode of 5, and not acknowledged.

Acknowledgement

  • ACK packet is used to acknowledge selected received EIGRP packets to facilitate their reliable delivery.
  • Sent in response to Update, Query, Reply, SIA-Query, and SIA-Reply.
  • Unicasted to the sender of the acknowledged packet.
  • ACK is essentially a Hello packet with an empty body (no TLVs), and having a non-zero Acknowledgement number field who value is set to the Sequence number of the reliable packet being acknowledged.
  • The same Opcode as the Hello packet, i.e. 5.
  • If both an ACK and a reliable packet are waiting to be sent to the same neighbor, EIGRP can put the acknowledgment number into the reliable packet’s Acknowledgment number field, saving the need to send a standalone ACK.
  • Only unicast packets can also be used to carry acknowledgement numbers.

Update

  • Contains routing information updates and is used to convey the reachability of destinations.
  • Can be both unicasted and multicasted.
  • Rules:
  1. During a new adjacency buildup, Update packets are unicasted between the newly discovered neighbors. In specific cases, like DMVPN, EIGRP might choose to synchronize neighbors using multicasts. The choice only impacts the efficiency of the initial synchronization process, and has no influence on the actual contents of exchanged information.
  2. After routers have fully synchronized, further Updates are sent as multicasts.
  3. If a neighbor does not acknowledge the arrival of an Update packet, EIGRP will retransmit the Update as unicast to the unresponsive neighbor.
  4. On point-to-point interfaces and for statically configured neighbors, EIGRP always uses unicast.
  • Delivered reliably; always acknowledged and retransmitted if no acknowledgement is heard.
  • Opcode of 1.

Query

  • Used to involve neighbors in the task of searching for the best route toward a destination.
  • Delivered reliably.
  • Both unicasted and multicasted.
  • By default, on multiaccess interfaces with only dynamic neighbors, Queries are sent as multicasts.
  • If not acknowledged in proper time, a Query is retransmitted to the unresponsive neighbor as unicast.
  • On point-to-point interfaces and toward statically configured neighbors, Queries are always sent as unicasts.
  • While each received Query must be acknowledged with an ACK, this ACk does not constitute a response to the Query message, only an acknowledgement that the Query has been received.
  • The Reply is used for that purpose.
  • Opcode of 3.

Reply

  • Sent in response to Query packets and carry their sender's current distance to the destination after taking into account the topology change that prompted the Query.
  • Always unicasted to the originator of the Query message, and delivered reliably.
  • Opcode of 4.

SIA-Query and SIA-Reply

  • Used during a prolonged diffusing computation to verify whether a neighbor that has not yet sent a Reply to a Query is truly reachable and still engaged in the corresponding diffusing computation.
  • The SIA-Query is used to ask a neighbor to confirm that it is still working on the original Query.
  • If the neighbor is reachable and still engaged in the diffusing computation, it will immediately respond with an SIA-Reply.
  • The maximum diffusing computation timer is reset.
  • Both SIA-Query and SIA-Reply are unicast, and reliably delivered.
  • SIA-Query uses an Opcode of 10.
  • SIA-Reply uses an Opcode of 11.

Reliable Transport Protocol

  • Manages the delivery and reception of EIGRP packets.
  • Reliably delivered packets: Update, Query, Reply, SIA-Query, and SIA-Reply (both unicast and multicast).
  • A non-zero Sequence number is carried in the packet header.
  • The Sequence number is a global value maintained per each EIGRP process instance on a router and is incremented whenever one of these packets is originated by the instance, regardless of which EIGRP-enabled interface the packet is going to be sent from.
  • Hello and ACK packets (non-reliable) set their Sequence number to zero and do not cause the global Sequence number to increase.
  • Conditional Receive: EIGRP partitions its neighbors into two groups: 1) well-behaved neighbors that have been able to acknowledge all multicast messages sent so far, and 2) "lagging" routers that have failed to acknowledge at least one EIGRP packet.
  • For in-parallel operation, EIGRP has to send the in-order multicast packets with a special flag saying "this packet is only for those routers who have received all multicast packets so far".
  • Sequenced Hello with two specific TLVs: the Sequence TLV and the Next Multicast Sequence TLV.
  • The Next Multicast Sequence TLV contains the upcoming sequence number of the next reliable multicasted message.
  • The Sequence TLV contains a list of all lagging neighbors by their IP address, in effect saying "whoever finds himself in this list, ignore the next multicast message with the indicated sequence number".
  • A neighbor receiving the Sequenced Hello and not finding itself in the Sequence TLV will know that it is expected to receive the upcoming multicast packet, and will put itself into a so-called Conditional Receive mode (CR-mode).
  • The next multicast packet is sent with the CR flag set.
  • Routers in CR-mode will process this packet as usual and then exit the CR-mode; other routers will ignore it.
  • As a result, the router is able to continue using multicast with those routers that have no issues receiving and acknowledging it, while making sure that the lagging neighbors won't process the multicasts until they are able to cacth up.
  • The time to wait for an ACK before declaring a neighbor as lagging and switching from multicast to unicast is specified by the multicast flow timer.
  • The time between the subsequent unicasts is specified by the retransmission timeout (RTO).
  • Both the multicast flow timer and the RTO are calculated for each neighbor from the smooth round-trip time (SRTT).
  • The SRTT is the average elapsed time in milliseconds between the transmission of a reliable packet to the neighbor and the receipt of an acknowledgement.

Router Adjacencies

  • EIGRP routers establish and maintain neighbor adjacencies.
  • By default, EIGRP discovers neighbors dynamically but neighbors can also be manually defined.
  • Dynamic neighbor discovery is performed by sending EIGRP Hellos to 224.0.0.10 or FF02::A as soon as as EIGRP is activated on an interface.
  • Static neighbors are typically deployed across media that does not natively support broadcast or multicast packets, such as Frame Relay.
  • After a static neighbor is defined, all EIGRP multicasts on the interface, through which the neighbor is reachable, will be disabled.
  • EIGRP neighbors must agree on the following parameters:
  1. EIGRP authentication parameters (if configured)
  2. EIGRP K values
  3. EIGRP Autonomous System (AS) number
  4. use of primary addresses for EIGRP neighbor relationships
  5. use of the common IP network address on a single subnet
  • If two routers differ in any of these parameters, they will not become EIGRP neighbors.
  • The Hello and Hold timers do not need to match between the neighbors.
  • The Hold time is the maximum time a router should wait to receive subsequent valid EIGRP packets from a neighbor.
  • If the Hold timer expires, the neighbor is declared unreachable and DUAL is informed.
  • By default, the Hold time is three times the Hello time, equaling either 15 or 180 seconds, depending on the interface type.
  • Changing the Hello interval does not result in automatic recalculation of the Hold time.
  • The timers can be changed at the interface level: ip hello-interval eigrp as-num sec and ip hold-time eigrp as-num sec.

EIGRP Neighbor Table

  • Records information about each detected neighbor with whom this router has established an adjacency.
  • The H (Handle) number is used to internally identify neighbors in an address-family independent way.
  • The Address and Interface columns hold the neighbor's IP address and this router's interface toward the neighbor.
  • The Hold time is derived from the value advertised by the neighbor and decremented each second; reset every time an acceptable EIGRP packet from that neighbor is received.
  • The Uptime shows the time the neighbor has been up.
  • The SRRT estimates the turnover time between sending a reliable packet to the neighbor and receiving an appropriate acknowledgement.
  • The RTO is the time that the router will wait for an acknowledgement of a retransmitted unicast packet before sending another copy.
  • SRRT and RTO are shown in milliseconds.
  • The Q Cnt indicates the number of enqueued reliable packets for which no ACK has been received yet.
  • In a stable network, the Q Cnt value must be zero.
  • Non-zero values are normal during initial router database synchronization or during network convergence.
  • The Sequence number shows the sequence number of the last reliable packet received from a neighbor.
  • The Sequence number is a per-process variable.
  • It is normal to see the sequence number increase by more than 1, as long as the value is ascending.

Tuesday, January 13, 2015

CCIE OCG: EIGRP (Chapter 8), Part I

EIGRP Basics and Evolution

Basic EIGRP is no longer a closed, proprietary protocol. In 2013, Cisco opened up the EIGRP specification and published it as an IETF Internet Draft, a precursor to an RFC.

EIGRP Feature Summary

  • Transport: IP, protcol type 88 (does not use UDP or TCP). Implements its own Reliable Transport Protocol, providing reliable unicast and multicast packet delivery.
  • Metric: Based on constrained bandwidth and cumulative delay by default, and optionally load and reliability.
  • Hello interval: Interval at which routers send EIGRP Hello messages on an interface.
  • Hold timer: Timer used to determine when a neighboring router has failed, based on a router not receiving any EIGRP messages, including Hellos, in this timer period.
  • Update destination address: Normally sent to 224.0.0.10 or FF02::A, with retransmissions being sent to each other's unicast IP address.
  • Full or partial updates: Full updates are used when new neighbors are discovered; otherwise, partial updates are used.
  • Authentication: Supports MD5 and SHA-based authentication.
  • VLMS/Classless: EIGRP includes the mask with each route, also allowing it to support discontigous networks and VLSM.
  • Route Tags: Enables EIGRP to tag and filter internal and external routes using distribute-lists and route-maps.
  • Next-hop field: Supports the advertisement of routes with a different next-hop router than the advertising router.
  • Manual route summarization: Allows route summarization at any point in the EIGRP network.
  • Multiprotocol: Supports the advertisement of IPv4 and IPv6. Former implementations also supported IPX and AppleTalk routes.

EIGRP Roots: Interior Gateway Routing Protocol

  • IGRP is a dead protocol that was an immediate predecessor to EIGRP.
  • IGRP was developed to eliminate RIPv1's working but naive hop count metric and the network diameter limit of 15.
  • IGRP relied on a composite metric made up of a variety of route variables.
Benefits of IGRP over RIPv1:
  1. wider network diameter, up to 255 hops
  2. complex multivariate metric
  3. unequal-cost load sharing
  4. an update period of 90 seconds, three times longer than RIP
  5. a more efficient update packet format
  • IGRP was designed to interoperate with multiple routed protocols, including IPv4, IPX, and AppleTalk.
  • IGRP broadcasted a Request packet out all IGRP-enabled interfaces at startup and performed a sanity check on received Update packets to verify that the source address of the packet belonged to the same subnet on which the Update was received.
  • Update packets were sent periodically every 90 seconds.
  • Like RIPv1, a classful distance-vector protocol.
  • Also relied on Split Horizon, triggered updates, Invalid after, Holddown, and Flushed after timers for functional stability.
  • IGRP summarized advertised addresses at network boundaries.
  • Overall, IGRP was better than RIP for larger networks, but it still had many of the fundamental limitations of RIP.

Moving From IGRP to EIGRP

Most significant IGRP weaknesses and detractors of IGRP include:
  1. sending full routing updates periodically
  2. the lack of VLSM support
  3. slow convergence
  4. the lack of adequate loop-prevention mechanisms
  • There is, in fact, very little left of the original IGRP in EIGRP - even though EIGRP is still a distance-vector routing protocol, albeit an advanced one.
  • In IGRP and RIP, the periodic Updates provided two separate and unrelated functions: 1) detecting neighbors and their liveliness, and 2) carrying routing information.
  • EIGRP decouples the building and maintaining of adjacencies from exchanging routing information.
  • The first task is accomplished by using a Hello protocol.
  • It is no longer necessary to periodically send Update packets to announce a router's continuous presence on a network. The Hello protocol is used to build and maintain neighbor adjacencies.
  • To send triggered updates, EIGRP implements the Reliable Transport Protocol (RTP; do not confuse with Real-time Transport Protocol), which is a Layer 3-independent transport protocol capable of reliable unicast and multicast delivery.
  • The use of RTP allows routers to initially exchange the complete routing information when synchronizing for the first time, and afterward, advertise changed routes only.
  • EIGRP completely abandons the periodic updating process and operates in an event-driven, incremental fashion.
  • To maintain a loop-free operation at every instant, EIGRP uses a so-called Feasibility Condition. 
  • So-called diffusing computations work in tandem with the Feasibility Condition. 
  • The current least-cost path can become ineligible, and EIGRP queries its neighbors to update their own best path selection in relation to the topology change and reply with their updated distances. Neighbors that are not adversely affected will simply respond right away; neighbors that are adversely affected will propagate the query further and can respond only after receiving all replies and making a choice themselves.
  • EIGRP implements a finite state machine called Diffusing Update Algorithm (DUAL) that controls the run of a diffusing computation, processing the replies and eventually inserting the gathered information into the routing table or commencing an additional computation.
  • The DUAL is not the diffusing computation algorithm itself; it is a control mechanism on top of diffusing computations.
  • EIGRP has a default hop-count limitation of 100 (can be manually adjusted with metric maximum-hops <1-255>).
  • EIGRP makes a distinction between internal and external routes.
  • Internal routes are those injected into EIGRP by the network command (AD of 90 by default).
  • External routes are redistributed into EIGRP from a different routing source (AD of 170 by default).
  • Changing the default AD values:  distance eigrp int ext.

EIGRP Classic Metrics

  • EIGRP uses several metric components: bandwidth, delay, reliability, load, MTU, and hop count.
  • The first four are combined together using a well-known formula to produce a single number, i.e. the composite metric, which is used by EIGRP to choose the best path toward a destination.
  • EIGRP shipped with recent IOS versions also supports a so-called Wide Metrics, which expand the allowable range of existing Classic Metrics.

Bandwidth

  • A static metric assigned to each router interface using the bandwidth command.
  • Obviously, describes the transmission speed of an interface.
  • Expressed in kilobits per seconds.
  • If no bandwidth is specified, the IOS assigns an implicit bandwidth value to each interface, depending on its hardware type and operational characteristics.
  • With Ethernet interfaces, the implicit bandwidth value reflects the true speed negotiated by the interface with its links partner.
  • On other interface types, the value has no realistic relation to the interface capabilities.
  • EIGRP considers the minimum bandwidth along the route.
  • This is done by comparing the bandwidth as advertised by neighboring router to the bandwidth of the interface toward the advertising neighbor, and taking the lower value of these two.
  • With Classic Metrics, EIGRP is capable of describing the bandwidth in the range of 1 kbps up to 10 Gbps.

Delay

  • A static metric assigned to each interface using the delay command.
  • Estimates the serialization delay incurred by the interface.
  • Not related to any true characteristic of an interface.
  • Inconvenient units: tens of microseconds.
  • Configuring a delay of 123 defines the delay to be 1230 microseconds.
  • The show interface command output reports the interface delay directly in microseconds.
  • Therefore, the show interface output will always display a value ten times higher than the configured value.
  • If the delay is not configured explicitly, IOS assigns an implicit delay value to each interface, depending on the interface hardware type.
  • When calculating the composite metric, EIGRP takes the total delay into account.
  • This is done by taking the delay as advertised by a neighboring router and summing it with the delay of the interface toward the advertising neighbor.
  • With Classic Metrics, EIGRP is capable of describing the delay in the range of 10 to 167,772,140 microsends.
  • The maximum value is used to indicate an infinite distance.
  • Split Horizon with Poisoned Reverse, Route Poisoning, withdrawing a route - all these techniques in EIGRP use the maximum delay to indicate an unreachable route.

Reliability

  • A dynamically estimated metric of an interface that evaluates its ratio between successfully received and all received packets.
  • The ratio is expressed as a fraction of 255.
  • A reliability of 255 expresses a 100 percent reliability.
  • EIGRP takes the minimal reliability into account.
  • This is done by comparing the reliability as advertised by a neighboring router to the reliability of the interface toward the advertising neighbor, and taking the lower value of these two.
  • A change in interface reliability value does not trigger sending EIGRP updates.
  • Currently just a relic carried over from its predecessor, with no particular usability.

Load

  • A dynamically estimated metric that measures the amount of traffic flowing through the interface in relation to its maximum capacity.
  • Also expressed as a fraction of 255, with the load of 1 representing an emptry interface and the load of 255 representing a fully utilized interface.
  • IOS actually computes an exponentially weighted average over the momentary load that smooths out short-lived load swings.
  • Two independent load metric counters: the Txload for outgoing traffic and Rxload for incoming traffic.
  • EIGRP takes the maximal Txload into account.
  • This is done by comparing the load as advertised by a neighboring router to the Txload of the interface toward the advertising neighbor, and taking the higher value of these two.
  • The changes in the Txload values on interfaces do not trigger EIGRP updates.
  • Also a relic from IGRP with no particular usability in EIGRP.

MTU

  • EIGRP advertises the minimum MTU along the route to the destination.
  • However, the MTU is completely unused in the best-path selection process.
  • Not factored into the composite metric, nor is it used as any kind of a tiebreaker.

Hop Count

  • Simply a counter of routers (hops) in the path toward the destination.
  • It is just a fallback security measure: routes reaching over a predefined threshold are advertised as unreachable, thereby breaking any potential routing loops.
  • By default, the limit is 100, configurable in the range of 1 to 255.
  • Not factored into the composite metric calculation and does not impact the best-path selection in any way.

Calculating the Composite Metric

  • Because EIGRP treats each metric component differently (delay is summed, bandwidth and reliability are minimized, load is maximized), EIGRP routers exchange these component metrics as separate values.
  • Each router independently computes the resulting composite metric on its own.
  • This composite metric is used locally on a router, and is never advertised as a single number in EIGRP messages.
  • The only exception to this rule is when a route is redistributed from one EIGRP process to another.
  • Even in this case, the composite metric of the redistributed route is retaken only for diagnostic purposes and carried separately from the seed metric.
  • EIGRP computes the composite metric value using the following formula.

  • The constants K1 through K5, commonly called K values, are weight constants in the range 0-255 that can be tweaked to influence the impact of individual metric components on the overall composite metric.
  • It is crucial that all EIGRP routers in the same domain compute the composite metric in the same way.
  • Therefore, K values on all routers must match.
  • If K values do not match, two routers will not be able to establish an adjacency.
  • By default, K1 and K3 are set to 1, and all other K values are set 0, causing EIGRP to take only bandwidth and delay into account.
  • The multiplication by 256 is the result of expanding the former IGRP 24-bit metric into a 32-bit metric used by EIGRP.

EIGRP Wide Metrics

  • EIGRP Classic Metric faced issues with interfaces faster than 1 Gbps.
  • The Bandwidth component is unable to differentiate between 10 Gbps and faster interfaces.
  • The Delay component is already set to the lowest value of 1 (microseconds) on 1 Gbps interfaces, and cannot be made smaller.
  • Cisco routers perform integer arithmetic.
  • To verify support of Wide Metric, issue show eigrp tech-support and look for Wide Metrics in the output.
  • show ip protocols: the presence of the K6 constant, the rib-scale of 128 and 64-bit metric version all indicate that Wide Metrics are supported.

EIGRP Wide Metrics consist of the following components:
  • Throughput: Analogous to the classic Bandwidth component. Calculation: 65536x10^7/Interface Bandwidth in kbps. Effectively tells how many times slower the interface is than a 655.36 Tbps link.
  • Latency: Analogous to the classic Delay component. Calculation: 65536xInterface Delay/10^6. The interface delay is expressed in picoseconds. The computation of per-interface delay metric differs based on its physical capability and configuration.
  • If < 1 Gbps without bandwidth and delay commands, the interface delay is simply its IOS-based default delay converted to picoseconds.
  • If > 1 Gbps without bandwidth and delay commands, the interface delay is computed as 10^13/interface default bandwidth.
  • With explicit bandwidth but without delay, regardless of their physical operating speed, the interface delay is the IOS-based default delay converted to picoseconds.
  • With explicit delay, regardless of the physical operating speed and the bandwidth setting, the interface delay is computed as its specified delay value converted to picoseconds.
  • Reliability: Identical to the classic Reliability component, and has not changed in Wide Metrics.
  • Load: Identical to the classic Load component, and has not changed in Wide Metrics.
  • MTU: Identical to the classic MTU, and is advertised but not used.
  • Hop Count: Identical to the classic Hop Count, and is advertised but unused in path selection; it only prevents potential routing loops.
  • Extended Metrics: Considered as placeholders for future extensions to the composite metric computation. Three extended metrics are defined: Jitter, Energy, and Quiescent Engergy. K6 was introduced to incorporate these metric components. Not usually used or supported.

Wide Metric formula.


  • The Wide Metric support is available when EIGRP is configured in named mode, and is automatically activated (no command to control the activation).
  • EIGRP automatically detects if their neighbors also support Wide Metrics, and use the appropriate metric type when talking to them.
  • Wide Metrics are preferred.
  • In the case of mixed neighbors on a common interface, EIGRP will use both metric formats, allowing each neighbor to process the metric format it supports.
  • The Wide Metrics composite value can result in a number wider than 32 bits, which is the maximum supported by the the RIB.
  • The value must be downscaled before the route can be passed down to the RIB.
  • This is done by dividing the value by a factor configured in the metric rib-scale command.
  • The default value is 128 and can be configured in the range 1-255.
  • The downscaled value is not used by EIGRP in any way.
  • EIGRP makes all its path selections based on the Wide Metrics composite value; only after a best path toward a destination is selected, its composite metric value is downscaled as the route is installed to the RIB.

Tweaking Interface Metrics to Influence Path Selection

  • Only Bandwidth and Delay can be manually influenced.
  • Do not use Bandwidth for traffic engineering. The change will only affect path selection if the configured value is the lowest bandwidth over the entire path. Further, changing the bandwidth can have an impact beyond affecting the EIGRP metrics (for example, QoS).
  • By default, EIGRP throttles to use 50 percent of the configured bandwidth. Lowering the value can starve EIGRP neighbors from getting packets. An excessively high bandwidth can lead EIGRP to consume more bandwidth than physically available, leading to packet drops.
  • The delay does not impact other protocols nor does it cause EIGRP to throttle back, and as the sum of all delays, it has a direct effect on path selection.
  • Bandwidth should always be configured to the true bandwidth of the interface and should never be used to influence EIGRP's path selction.

Friday, January 9, 2015

CCIE OCG: RIPv2 and RIPng (Chapter 7)

Introduction to Dynamic Routing

  • Constitute messages exchanged between routers to detect their mutual presence and convey information about the existing networks in the topology, and algorithms necessary to process this information and advertise it throughout the network.
  • Distance-vector routing protocols are principally founded on the exchange of, well, distance vectors, or arrays of distances to known networks. 
  • The term vector = unidimensional array.
  • The term distance = the measure of feasibility, or a metric, of reaching a particular network.
  • Messages exchanges contain arrays, with each element containing information about networks known to the router originating the message, and that router’s distance to these networks.
  • A router learns about the existence of a network by receiving a message from its neighboring router that advertises the network. 
  • This neighbor then becomes the next hop toward this network.
  • After, the router starts advertising the route itself, announcing its own distance from the destination.
  • Each router advertises its own directly connected networks added to the routing protocol, and all routes learned by that protocol that have been placed into the routing table.
  • A router advertises only those routes used by itself.
  • Distance vector routing protocols exchange only lists of known networks and their distances.
  • They do not exchange information about the network's topology.
  • Generally prone to the creation of routing loops, but various avoidance methods exist.
  • Path-vector routing is an extension of distance-vector principles.
  • In addition to exchanging messages about known networks and their distances, each network is accompanied by a list of path elements.
  • These path elements can be theoretically anything.
  • BGP is the only common path-vector routing protocol.
  • In BGP, autonomous system numbers guarantee a loop-free topology.
  • A router will not accept an advertisement about a network whose path description already includes the identifier the router would put into that advertisement itself.
  • Link-state routing protocols exchange information about individual objects in the topology and their interconnections. 
  • IP prefixes are often treated only as attributes, or properties, of these objects.
  • Every router has exact information about the entire area’s topology: It knows every router, every network, and every link. This detailed information about each object in an area then allows the router to construct a so-called directed graph of the topology (in essence a map of the topology) and compute a tree of shortest paths to find all reachable destinations and least-cost paths toward them. 
  • Usually, the Shortest Path First (SPF) algorithm invented by Edsger W. Dijkstra is used.
  • Dijkstra’s algorithm by its very nature does not construct routes containing loops.
  • Temporary routing loops can still occur when routers do not compute shortest paths using the same information - usually during network convergence.
  • Link-state downside: inability to perform route summarization, filtering, or applying offset-lists in arbitrary parts of the network. 

RIPv2 Basics

  • General characteristic: classless, distance-vector, timer-driven routing protocol.
  • Transport protocol: UDP, port 520.
  • Metric: Hop count, with 15 as the maximum usable metric, and 16 considered to be infinite.
  • Hello interval: none; RIPv2 relies on the regular full routing updates instead.
  • Update destination: 224.0.0.9.
  • Update interval: 30 seconds.
  • Full or partial updates: Full updates at each interval. For on-demand circuits, allows RIPv2 to send full updates once, and then remain silent until changes occur (RFC 2091).
  •  Triggered updates: Yes, when routes change.
  • Authentication: plain-text and MD5.
  • Route tags: Yes, as routes are redistributed into RIPv2.
  • Next Hop field: Supports the assignment of a next-hop IP address for a route, allowing a router to advertise a next-hop router that is different from itself.

  • RIPv2 exchanges routes by sending RIPv2 updates on each RIPv2-enabled interface based on the Update timer.
  • RIPv2 router advertises its connected routes, and RIPv2-learned routes if they are in the  routing table.
  • RIPv2 routers do not form neighbor relationships, nor do they use a Hello protocol.
  • Each router simply sends updates to 224.0.0.9.
  • RIPv2 routers can also be configured to use 255.255.255.255 using the ip rip v2-broadcast per-interface command (not commonly done).
  • Two types of messages: Request and Response.
  • The message format for both message types is identical.


  • Command field is set to 1 for Request and 2 for Response.
  • Each routing entry occupies 20 octets in total.
  • At most 25 routing entries can be placed into a single RIP message.
  • A Request message is used to ask a neighbor to send a partial or a full RIP update immediately, rather than wait for the Update timer to expire.
  • Full RIP update Request contains exactly one routing entry with the address family ID set to 0 and metric set to 16.
  • If a Request message lists one or more networks, only the update on these networks is requested.
  • A full RIP update is requested when the RIP process is being started, a RIP-enabled interface comes up, or when the clear ip route * command is issued.
  • RIPv2 uses hop count as metric.
  • 15 is the largest valid metric, and 16 is considered to be infinity.
  • RIPv2 increments the metric when sending updates (RIPng and EIGRP increment their metrics when receiving updates).
  • By default, RIPv2 installs 4 equal-cost paths into the routing table, as dictated by the maximum-paths command (1-16).
  • The actual upper limit depends on the IOS version and router platform, and is not in fact related to RIPv2 or any other protocol.

RIPv2 Convergence and Loop Prevention

  • Counting to Infinity: If the next-hop to a network suddenly advertises that network with an increased metric, accept the advertisement immediately and update our metric accordingly. If the updated metric reaches infinity, stop using that next hop.
  • Split Horizon: Do not advertise routes whose outgoing interface field matches the interface out which the update would be sent.
  • Split Horizon with Poison Reverse: A stronger variant of Split Horizon; All routes whose outgoing interface matches the interface out which the update would be sent are advertised with an infinite metric.
  • Route Poisoning: Sending an infinite-metric route in routing updates when that route fails.
  • Triggered update: The immediate sending of a new update when routing information changes, instead of waiting for the Update timer to expire. Only the change network is sent in the triggered update. Complete updates continue to be sent in regular intevals.
  • Update timer: Specifies the time interval over which updates are sent, 30 seconds by default.
  • Invalid after timer: A per-route timer (default 180 seconds) that is reset and begins after an update about a route has been received from its next hop. If the updates about the route from its next hop cease to be received and the Invalid after tiemr reaches its limit, the route is declared invalid and the Holddown timer starts for this route.
  • Holddown timer: A per-route timer (default 180 seconds) that begins after a route has been declared invalid (the Invalid after timer has expired). The router starts advertising that route as unreachable, does not accept any updated information, and does not modify the routing table entry for that route until the Holddown timer for that route expires.
  • Flushed after timer: A per-route timer (default 240 seconds) that is reset and begins after an update about a route has been received from its next hop. If the updates about the route from its next hop cease to be received and the Flushed after timer reaches its limit, the router removes the route from the routing table entirely.

  • Routers exchange lists exchange lists, or vectors, of known networks and their distances.
  • For each network, a router chooses the neighbor providing the least total metric towards the network.
  • All routers advertising the same network with a total higher cost are ignored. 
  • One exception: if the next-hop router for a network suddenly advertises a higher distance than last time, this advertisement is not ignored but rather accepted immediately.
  • The router will update the total metric and advertise it. The network and the current next-
    hop are kept in the routing table.
  • Only a subsequent arrival of an update from a different neighbor providing a lower total
    metric would cause the router to change the next hop.
  • Count to Infinity: As per the logic of immediately accepting a higher metric from a next-hop, two routers will increase the metric of a route to infinity with a loop prevention mechanism.
  • The concept of a finite metric, i.e. a metric value representing an unreachable network, will eventually break the loop and drop the route from the routing table.
  • Split Horizon: a network should never be advertised back over the interface that is used to reach that network.
  • By default, Split Horizon is activated on most interfaces, notable exceptions being physical Frame Relay and ATM interfaces.
  • Split Horizon with Poison Reverse: a network should always be explicitly advertised as unreachable over the interface that is used to reach that network.
  • While more effective than merely Split Horizon, it is not implemented in the Cisco RIPv2.
  • Route Poisoning is a mechanism used to rapidly flush a route that has become unreachable.
  • Accomplished by advertising this route with the metric set to infinity.
  • A router that receives an update about a network from its next-hop with an infinite metric will immediately remove the route from the routing table.
  • The router will itself advertise that network as unreachable.
  • Routers that receive an update about an unreachable network from different neighbors than their respective next-hops are not influenced.
  • The message is indistinguishable from a Split Horizon with Poisoned Reverse advertisement. 
  • They process it accordingly but it has no effect on their routing tables.
  • Even though a route is removed from the routing table during a Route Poisoning procedure, RIP will still keep the route in its internal database marked as possibly down (show ip rip database) to allow repeated advertisements because a single advertisement can get lost (UPD-based protocol, no acknowledgements).
  • The unreachable route will be flushed from the RIP internal database after Flushed after - Invalid after seconds.
  • It is important to distinguish between Poisoned Reverse and Route Poisoning.
  • Both refer to advertising a network with an infinite metric.
  • Poisoned Reverse is an extension to the Split Horizon principle that advertises a route as unreachable back to its next-hop to prevent it from creating a routing loop.
  • Route Poisoning advertises a truly unreachable route to quickly flush it from the routing table to allow a backup path to take over.
  • Triggered updates are sent in the moment of detecting a change in reachability of a network, rather than waiting for the full Update interval to expire.
  • The update commonly carries only the changed network, without listing all other known networks.
  • The other networks continue to be advertised, along with the changed information, in regular intervals driven by the Update timer.
  • Triggered updates are also called flash updates.
  • The Holddown mechanism: delay processing updates about a network whose reachability has become questionable, as the received updates might not yet contain up-to-date information.
    The on-going lack of any information about a network from its next-hop is tolerable for a limited time period, but it exceeds a reasonable time, it is clear that "something happened".
  • The router that has detected a sudden loss of reachability information for a network must not immediately accept the updates from its neighbors; rather, it must give them certain time to learn about the outage and converge to a different path. Only after this time, updates from other routers regarding the missing network can be accepted as trustworthy again.
  • Two independent RIPv2 timers: the Invalid after timer and Holddown timer.
  • The Invalid after time is reset every time an update is received.
  • A route is considered usable within the Invalid after timer (180 seconds by default).
  • If the Invalid after timer has reached its limit and an update has not been received, the network is considered invalid. In such case, the following happens:
    1. The router declares the network invalid. Visible in the show ip route output as "is possibly down". The Invalid after timer is stopped for this network.
    2. The router starts the Holddown timer for this network. While the timer runs (180 seconds by default), the router itself advertises the network with infinite metric (Route Poisoning) to force its neighbors to find an alternative route if possible.
    3. The router also locks the entry in the routing table. Absolutely no updates whatsoever are accepted until Holddown timer expires, regardless of who sends them and what metric they claim.
    4. After the Holddown timer expires, the router unlocks the routing entry in its routing table and converges through a neighbor that offers the lowest metric route to the network.
     
  • The procedure is not triggered by the arrival of an update that advertises an unreachable network (Route Poisoning). Rather, this procedure is invoked after the reachability of a network can be neither confirmed nor refuted for a period of time.
  • After a router puts a route into the invalid state, it advertises that route as unreachable itself. "I know there is some problem reaching this network, but I don't know exactly what. Whoever uses me as a next-hop, stop doing that."
  • A router never updates an invalid route while the Holddown timer runs.
  • If there is a detour path, the router will learn about it after the Holddown timer expires and the routing entry is unlocked. However, the failed route might have been the only path to the destination, and there might be no alternative path available, so no neighbor will advertise it. That would cause the routing entry to linger in the routing table indefinitely, still pointing toward the previous next hop.
  • To prevent a route that has stopped being advertised from lingering in routing tables indefinitely, yet another timer is present: the Flushed after timer.
  • Similarly to the Invalid after timer, the Flushed after timer is reset every time an update about a network is received.
  • If the Flushed after timer reaches its limit, the route is immediately flushed from the routing table, and advertised as unreachable.
  • Default timer values: Update - 30 seconds, Invalid after and Holddown - 180 seconds, Flushed after - 240 seconds.
  • The default timer settings actually does not allow the Holddown timer to completely expire.
  • If a network stops being advertised  by its next-hop, it will be put into an invalid state 180 seconds after the last update arrived (the Flushed after timer also starts). The Holddown timer then starts, but after 60 seconds, the Flushed after timer will expire, and the router will remove the route entirely along with all associated timers.
  • As a result, the effective Holddown period is only 60 seconds.
  • The Flushed after timer has no effect on a route that is still considered valid (that is, its age is less than the upper limit of Invalid after ).

RIPv2 Convergence When Routing Updates Cease

  • RIPv2 must wait for some timers to expire before routes, whose advertisements from the neighbor have stopped, can be considered failed.
  • RIPv2 uses its Invalid after, Flushed after, and Holddown timers to prevent loops.
  • RIPv2's convergence time increases to several minutes as a result.
  1. R3's interface fails, but R1's interface stays up, so R1 must use its timers to detect route failures.
  2. R1's Invalid after and Flushed after timers for a network grow because R1 does not hear any further updates from R3.
  3. After the Invalid after timer expires (180 seconds), R1 starts a Holddown timer for the route. Holddown starts at (default) 180 seconds and counts down.
  4. The Flushed after timer expires after 240 seconds, or 60 seconds past the Invalid after timer. As a result, R1 flushes the route from its routing table, which also removes the Holddown timer for the route.
  5. R1 receives R2's next regular full routing update, and place the route in its routing table, completing convergence.
  • Either the Flushed after timer or the Holddown timer must expire before new routing information would be used.
  • When the route is flushed (removed), any associated timers are also removed, including the Holddown timer.
  • Had the Holddown timer been smaller, and it expired before the Flushed after timer, R1 would have been able to use the route advertised by R2 at that point in time.

Convergence Extras

  • RIPv2 timers can be tuned with the timers basic <update> <invalid> <holddown> <flush> command under the RIP process.
  • The timers should be consistent across routers.
  • Smaller values increase the chance of transient routing loops being formed during convergence.
  • The clear ip route commadn bypasses loop-prevention features by deleting the route and timers. It can be risk, but it certainly speeds convergence.

RIPv2 Configuration

  • version 2: tells the router to send and receive only RIPv2 updates (RIPv1 updates are ignored).
  • network: always recomputes its argument (the network address) into a classful representation.
  • neighbor <ip-add>: send unicast RIP updates to that neighbor.
  • Enabling RIPv2 on an interface makes the router begin sending RIPv2 updates, listening for RIPv2 updates (UDP port 520), and advertising that interface’s connected subnet.

Disabling RIPv2 Functions

  • Sending RIPv2 updates: make the interface passive (passive-interface <int-num> under the RIP process.
  • Listening for RIPv2 updates: filter all incoming routes using a distribute-list, or filter incoming RIPv2 packets using a per-interface ACL.
  • Advertising the connected subnet: filter outbound advertisements on other interfaces using distribute-lists, filtering an interface's connected subnet.
  • If there is a multipoint interface with four routers and you want to send RIP updates to only one of them, make the interface passive, and then use the neighbor command.
  • RIPv2 uses autosummarization at classful network boundaries by default.
  • Automatic summarization occurs when a router intends to advertise a subnetwork of a particular classful network X (also called major network) out an interface that is itself in a different classful network Y. 
  • Disabling autosummarization should be considered a mandatory part of any RIPv2 configuration.

RIPv2 Authentication

  • Two requirements: creating the keys and enabling the feature on interfaces.
  • Clear-text or MD5 keys.
  • Multiple keys are allowed, and are grouped together using a construct called a key chain.
  • A key chain is simply a set of related keys, each identified with a different number and might be restricted to a time period.
  • NTP is recommended.
  • Authentication is enabled on a per-interface basis, referring to a key chain with ip rip authentication key-chain <name> command.
  • The route selects valid key(s) from the key chain.
  • If multiple keys are valid for signing outgoing RIPv2 packets, the key with the lowest sequence number will be used.
  • The type of authentication, clear-text or MD5, is chosen per interface as well (ip rip authentication mode {text|md5}).
  • If the type is not specified, the default is clear-text.
  • When authentication is enabled, the maximum number of prefixes that can be advertised in a RIPv2 message is reduced by 1 to a value of 24. The first route entry in each RIPv2 message would carry 20 bytes of authentication data.
  • With MD5, more authentication data would be placed after the entire RIPv2 message.
    1st route entry = MD5 authentication + 24 routing entries + more authentication date after the 25th entry.

RIPv2 Next-Hop Feature and Split Horizon

  • Split Horizon setting is controlled per interface (ip split-horizon).
  • On by default, except for Frame Relay and ATM physical interfaces.
  • The RIPv2 next-hop feature allows a RIPv2 router to advertise a different next-hop router than the advertising router.
  • On Cisco routers, this feature is not configurable, and is almost unused.

RIPv2 Offset Lists

  • RIPv2 offset-list allows RIPv2 to add to a route's metric, either before sending an update, or for routes received in an update.
  • The offset-list refers to an ACL, and the router adds the specified offset, or extra metric, to any matching routes.
  • Any routes not matched by the offset-list are unchanged.
  • The offset-list also specifies which routing updates to examine by referring to a direction (in and out) and, optionally, an interface.
  • If the interface is omitted, all updates for the defined direction are examined.

Route Filtering with Distribute Lists and Prefix Lists

  • Outbound and inbound RIPv2 updates can be filtered at any interface, or for the entire RIPv2 process.
  • The routes are filtered by a distribute-list, which references an access-list or a prefix-list, under the RIP process.
  • Any subnets matched with a permit clause make it through, deny actions are filtered.
  • The filtering can be performed for either direction and, optionally, for a particular interface.
  • If the interface is omitted, all updates coming into or out of the RIPv2 process are filtered.
  • Prefix lists are designed to match a range of subnets, as well as a range of subnet masks associated with the subnets.

RIPng for IPv6

  • RIPng is just a straightforward adaptation of RIPv2 for IPv6 operation with practically no changes to the underlying protocol mechanisms.
  • FF02::9, UDP 521.
  • Metric is again based on hop count: maximum of 15, 16 representing infinity.
  • RIPng differes in its metric handling: the metric is incremented by the receiver of a RIPng advertisement, not by the advertisement sender anymore.
  • All loop avoidance techniques described in previous sections are used by RIPng in precisely the same way.
  • The number of route entries in a RIPng message is limited only by the IPv6 MTU on the link, and the protocol itself poses no limitation on their count.
  • As RIPng does not assume any multiprotocol capability, the address family ID field has been omitted.
  • Because of the unpopularity of in RIPv2, the next-hop field would uselessly occupy a significant space (128 bits) in RIPng. Therefore, the next-hop field, if necessary, is specified by a separate route entry.
  • Authentication is not handled by RIPng anymore; these functions are offloaded to IPsec.

RIPng lacks several features supported in RIPv2:
  • Split Horizon can be activated or deactivated only on a per-process basis, not on individual interfaces.
  • Passive interfaces are not supported
  • Static (manual) neighbors cannot be configured.
  • Per-process offset-lists are not supported.
  • Still, there are management improvements over RIPv2:
  • Multiple RIPng process can be run on the router (up to 4). Individual processes are distinguished by an alphanumeric name that is local to the router and does not need to match between different routers.
  • Route Poisoning can be activated on a per-process basis.
  • Interfaces can be configured with a metric-offset value that is added to the metric in all received advetisements over that interface.
  • The default route can be originated on a per-interface basis, including an option of suppressing all other updates over that interface.

Saturday, December 20, 2014

Notes: CEF Polarization

Introduction

  • CEF polarization can cause suboptimal use of redundant paths to a destination network.
  • CEF polarization is in effect when a hash algorithm chooses a particular path and the redundant paths remain completely unused.

Background


Consider the following topology.


  • Assume that to reach host 10.1.1.1 from [Top Left] router, there are two equal-cost paths (L1, L2).
  • The decision of the chosen link is made by a hashing algorithm using the source and destination IP addresses (SIP and DIP).
  • When there are only two paths, the router performs an exclusive OR (XOR) operation on the lower-order bits of the SIP and DIP (one bit, two bits for 3-4 links, and so on).
  • The XOR of the same SIP and DIP always results in the same hash, and hence, the use of the same link.
  • At the next layer, the same hashing algorithm is used along with the same hash input, and a single link is picked for all flows, leaving the other link underutilized.
  • CEF polarization = use of the same hash algorithm and same hash input, which results in the use of a single ECMP link for ALL flows.


In more detail:

1. Traffic sourced from 10.240.18.1 and destined to 10.240.20.1 arrives at Router A, and the information goes through a hashing algorithm. The path toward Router C is chosen. From there the packets travel to Router F, and on to their final destination.

2. Router A performs the hashing function on traffic sourced from 10.240.18.2 and destined to 10.240.20.1, and sends the packets to Router B. The same happens to traffic from 10.240.18.3.

3. Router B produces the same results that the hash algorithm on Router A produced. Therefore, both streams of packets pass along one path towards Router E. The link toward Router D is never used.

4. The packets are forwarded to their final destination through Router F.


Avoiding CEF Polarization


1. Alternate between default (SIP and DIP) and full (SIP + DIP + Layer 4 ports) hashing inputs configuration at each layer of the network.

The Catalyst 6500 provides a few choices for the hashing algorithm:
  • Default - Use the source and destination IP address, with unequal weights given to each link to prevent polarization.
  • Simple - Use the source and destination IP address, with equal weight given to each link.
  • Full - Use source and destination IP address and Layer 4 port number, with unequal weights.
  • Full Simple - Use the source and destination IP address and Layer 4 port number, with equal weights given to each link.

  • No commands (except show run) exist to check the load-sharing algorithm in use.
  • The Catalyst 6500 does not support per-packet load sharing.
  • The Full option does NOT include a universal ID in hash.

2. Alternate between an even and odd number of ECMP links at each layer of the network.
CEF load balancing is independent of the route source. OPSF routes exhibit the same behavior EIGRP.

The hash algorithm load-balances this way by default:

1: 1
2: 7-8
3: 1-1-1
4: 1-1-1-2
5: 1-1-1-1-1
6: 1-2-2-2-2-2
7: 1-1-1-1-1-1-1
8: 1-1-1-2-2-2-2-2

This means that:
  • For two equal cost paths, load-sharing is 46.666%-53.333%, not 50%-50%.
  • For three equal cost paths, load-sharing is 33.33%-33.33%-33.33% (as expected).
  • For four equal cost paths, load-sharing is 20%-20%-20%-40% and not 25%-25%-25%-25%.

  • Evidently, when there is an even number of ECMP links, the traffic is NOT load balanced!
  • Anti-polarization weight was introduced in Version 12.2(17d)SXB: mls ip cef load-sharing full simple
  • Use this command if there are two equal-cost paths, and both need to be used equally.

3. A concept called unique-ID/universal-ID helps with CEF polarization.
  • The universal ID is a randomly generated value at router bootup that can be controlled manually.
  • A universal algorithm (the default in current IOS versions) adds a 32-bit value to the hash function.
  • This seeds the hash function on each router with a unique ID, which ensures that the same source/destination pair hash into a different value on different routers along the path.
  • Result: better load sharing and lowered polarization.
  • The unique-ID concept does not work for an even number of equal-cost paths due to hardware limitation.
  • To overcome this problem, Cisco IOS adds one link to the hardware adjacency table when there is an even number of equal-cost paths in order to make the system believe that there is an odd number of equal-cost links.
  • Customize the universal ID with ip cef load-sharing algorithm universal <id>.

References

CEF Polarization

Notes: Load Balancing with CEF

Introduction

  • IOS 11.1(17)CC introduced CEF for Cisco 7200 and 7500 series routers
  • CEF can efficiently use multiple parallel links without additional hardware multiplexers.
  • More information: "Alternatives for high bandwidth connections using parallel T1/E1 links."

Load Balancing

  • A functionality that distributes packets across multiple links based on routing information.
  • If multiple paths to the same destination are discovered, they are installed into the routing table according to the route selection process (discussed previously).
  • Only parallel links of the same metric are discussed here (although, the same principles apply over unequal cost paths).
  • Load balancing does not have to be supported. A router can use a single path, even if additional equal-cost paths exists (maximum-paths 1).
  • By default, a router will try to do load balancing.
  • The default in IOS is 4 entries for most IP routing protocols.
  • For BGP, the default is only one entry.
  • The maximum number that can be configured is 6 different paths.
  • Two modes of load balancing: per-destination and per-packet basis.

Per-Destination

  • All packets for a given destination are forwarded along the same path.
  • Preserves packet order.
  • Potential unequal usage of links.
  • A larger number of destination addresses leads to more equally used links.
  • A route-cache entry for every destination address vs. every destination network as done when only a single path exists.
  • Therefore, traffic for two different hosts on the same destination network can use different paths.
  • Core routers memory and processing requirements become demanding.

Per-Packet

  • Guarantees equal load across all links.
  • Packets may arrive out-of-order at the destination (due to differential delay in the network).
  • Disables the forwarding acceleration by a route cache, which includes the outgoing interface information.
  • Instead, determines the outgoing interface for each packet by looking up the routing table and picking the least used interface.
  • Ensures equal link utilization.
  • Process-intensive and impacts overall forwarding performance.
  • Not well-suited for higher speed interfaces.

Load Balancing with CEF for IP

  • Includes enhancements that allow to use load balancing without sacrificing forwarding performance even for per-packet load balancing, which previously required disabling of route caching mechanisms like fast switching or optimum switching.
  • Supports encapsulations such as Frame Relay, Ethernet, PPP, HDLC, and tunnels.

How CEF Load Balancing Works

  • CEF is an advanced Layer 3 switching technology inside a router.
  • The route cache is filled on demand when the first packet for a specific destination needs to be forwarded.
  • CEF takes a different approach as it calculates all information necessary in advance.
  • CEF also decouples the forwarding information from the next hop adjacency.
  • Two main CEF components are:
  1. Forwarding Information Base (FIB)
  2. Adjacency Table

 

Forwarding Information Base (FIB)

  • Used to make IP destination prefix-based switching decisions.
  • Conceptually similar to the routing table.
  • Maintains a mirror image of the forwarding information of the IP routing table.
  • When a routing or topology change occurs, the IP routing table is updated, and those changes are reflected in the FIB.
  • One-to-one correlation between FIB entries and routing table entries, so the need for route cache maintenance is eliminated.

Adjacency Table

  • Network nodes are adjacent if they can reach each other with a single hop.
  • CEF uses the adjacency table to prepend Layer 2 addressing information.
  • Maintains Layer 2 next-hop addresses for all FIB entries.
  • Populated as adjacencies are discovered.
  • Each time an adjacency entry is created (through the ARP protocol, for example), a link-layer header for that node is precomputed and stored in the adjacency table.
  • Once a route is determined, it points to a next hop and corresponding adjacency entry.
  • For each resolved path to a destination prefix, a pointer is added for the adjacency corresponding to the next-hop interface for that path.
  • For per-destination load balancing a hash is computed from the source and destination IP addresses.
  • This hash points to exactly one adjacency entry, providing the same path for all packets with this source/destination IP address pair.
  • For per-packet load balancing the packets are distributed in round-robin fashion over the available paths.
  • The FIB and adjacency tables provide all the necessary forwarding information.

Example:

Destination      Adjacencies    #    Interface    Encapsulation
192.168.25.0/24  34, 45         34   Serial0      PPP
                                45   Serial1      PPP


For each packet a match of the destination address against the FIB is made. In this example the FIB entry has multiple adjacencies for the destination 192.168.25.0, one of the adjacencies needs to be selected. For per-packet load balancing packets are encapsulated and forwarded by alternating between entry #34 and entry #45. For load balancing on IP address basis a hashcode is generated out of the addresses. For any given pair of source and destination addresses the hashcode is always the same, thus guaranteeing that always the same path is used.

CEF Configuration

  • Straightforward, as the only configurable option is the load balancing mode.
  • The global command to enable CEF is ip cef.
  • The command automatically enables CEF on all interfaces that use supported encapsulations.
  • The default load balancing mode: per-destination.
  • To change the load balancing mode from per-destination to per-packet use the interface command ip load-sharing per-packet.
  • Enabling CEF is sufficient in most cases.
  • Per-packet load balancing is only rarely needed, e.g. when the majority of traffic is between two hosts.

Summary

  • CEF supports load balancing for TCP/IP over parallel links without impacting performance even if the traffic patterns require per-packet load balancing.

References

Load Balancing with CEF