Showing results for 
Search instead for 
Did you mean: 
Cisco Employee
Cisco Employee



Broadband deployments are complex because of the options you have, varying needs in terms of deployments, the combination of technologies and many other reasons. In this article we'll go over the various designs, options and service deliverables that you can achieve with the ASR9000 BNG solution.


Access models

One of the decissions to be made when running BNG is the type of access that is preferred. There are 2 key options for the ASR9000 which is PPPoE (PPP over Ethernet) or IP sessions. Both can run on single or double tagged subinterfaces.


PPPoE sessions are triggered by the reception of a PADI and IP sessions are created by using DHCP as a session trigger.

In XR4.2.1 we can also use the "packet trigger" which means that an unclassified L3 address can be used to start a new session.


Also one has to decide on the Access interface, whether that is a single physical interface or access via bundles with multiple members. When choosing bundles the next decission is whether the members run in an active/active mode (that is all members are forwarding traffic) or in an active standby mode (whereby there is one link not forwarding traffic and only taking over when one of the member(s) fails).


Bundles vs Phyiscal interfaces


In the ASR9000 the use of a phyiscal interface, whether that be a GigabitEthernet or a TenGigabitEthernet interface are terminated on the linecard, this significantly increases the scale as effectively every linecard becomes its own controller of the sessions.

When you are running bundles, the sessions are then maintained on the RSP.


The LC's currently have their own Control Policy Engine, PPP manager and AAA processes. But they lack IGMP and LI today. This means that if you are planning to use LI or IGMP or parameterized QOS LC based subscribers should not be the choice for you.


You can use bundle interfaces with one member and disabling LACP to pull the subs to the RSP so that you have access to these features:


interface GigabitEthernet0/0/0/0

bundle id 100 mode on


When you are using bundle interfaces, all features enabled on the subscriber are programmed to the NP's where the bundle has members on.

So running a bundle with 2 members effectively creates the subscriber on both NPU's for those members, because of the phenomenal failover support it has.

For downstream loadbalancing we can use destination based hashing, which is the subscriber's ip address, so we always hash the traffic from one subscriber over 1 member, but the subscribers would be spread over the members based on their destination address.


The following configuration takes care of that:

interface Bundle-Ether100

bundle load-balancing hash dst-ip


When one member fails, the traffic is carried over to the other member seemlessly.

For the upstream direction, we can't control how the traffic arrives to us, which is controlled by the access layer device, whether that be your metro device or DSLAM.


Linecard based subscribers

This is an XR 5.1.1 deliverable. LC subs will increase the scale even further as it distributed the control plane to the LC as opposed to the RSP.


Hardware requirements

The hardware you require for BNG on the ASR9000 is:

  • Any typhoon linecard such as the A9K-MOD80-SE, A9K-24x10, 36x10 and MOD160.
  • A9K-RSP440-SE


Note that for both the RSP and the linecard we need to have the "SE" or Service Edge variant.

Also the ASR9001 will support BNG.


Trident linecards are not supported for subscriber termination, but they can be used as core facing linecard for transport


The MOD80 is a 2NP based linecard. Per NP we support 32k sessions with 64k per linecard.

Per Port we support up to 8K sessions when they are running QOS. This because of the QOS architecture within the NP that won't allow more then 8k parent shapers per phyiscal port. (that restriction is partially lifted in XR4.3.1 which allows for QOS chunk allocation more dynamically. You still have 32k parent shapers, but you can instruct a subinterface to use a chunk. This means a limitation of 8k per vlan, 32k per NPU)


These retrictions apply regardless of whether you are using LC or RSP based subscribers.


All linecards that run BNG need to be of the Typhoon kind (NP4). Your core facing interfaces if not running BNG can be Trident based or SIP700 based. However if you are using L2TP, then your core facing LC's need to be NP4 also. This is because the L2TP decap is not implemented in SIP or Trident which is handled by the core facing LC in the downstream direction.



Radius Source Ports

Each LC is assigned a source port range for radius, so while all LC's present themselves as a single NAS-ID, the source ports they are using effectively identify the node within the system. This is transparent for your use and nothing needs to be configured for that.

However it is imperative that your radius server supports extended source-ports (so the combination of source-port + radius-request-ID) defines a unique request. This is the same in IOS otherwise you can only support a call window of 256 (as the radius ID is only an 8 bit field). So check your radius-server capability!


For example see this radius request received on a radius-server:

Tue Apr 5 16:55:45 2011: [17224] message received from code=4, length=361

Tue Apr 5 16:55:45 2011: [17224] Acct-Interim-Interval = 60

Tue Apr 5 16:55:45 2011: [17224] Acct-Status-Type = Stop

Tue Apr 5 16:55:45 2011: [17224] Cisco-avpair = "if-handle=167774432"

Tue Apr 5 16:55:45 2011: [17224] Cisco-avpair = "client-mac-address=0010.9441.0001"

Tue Apr 5 16:55:45 2011: [17224] Acct-Session-Id = "00000054“


This radius (accounting) request was sent from UDP Port 49080 and its radiusID is number 12

You can check in LPTS in IOS-XR where that source port is mapping to:


RP/0/RSP1/CPU0:A9K-BNG#show lpts pifib hardware entry br location 0/1/CPU0 | i 49080

Tue Apr 5 17:22:52.941 EST

(VRF:0) any.49080 , any.any         UDP   any         48           deliver 48



If you decode 48 into binary: 48 =======> 0 0 0 0 1 1 0 0 0 0 (Binary of 48)

The "1" represent the RSP's in the middle slot of a 10 slot chassis!


PPPoE vs DHCP/IP-sessions

Broadband access has natively been using PPP based access, which originates from the dial days, whereby modems dialing into a modem-bank/access server allowing the transmission of data packets by encapsulating them in PPP packets. When access evolved to higher speeds using DSL (effectively ATM over the phone line), PPP was still used in the flavor of PPPoA (PPP over ATM).

Now that the aggregation point's uplinks are transitioning from ATM based to Ethernet based and the fact that there is Ethernet directly to the home, PPPoE has made a strong hold in the access layer.

Still with DSL in the first mile, DSLAM's may convert the PPPoA session into PPPoE towards the aggregator leaving PPP with a (well deserved) strong precense in the access.


In a transition to an all ethernet, there is no need per-se to run PPP at the access. PPPoE requires a client (which nowadays come natively in many operating systems however), which created the opportunity for a more simple approach of direct IP access using DHCP as a signal to trigger the session creation.


Drrivers of PPPoE are:

  • Ability to run multistack, each Layer 3 protocol is negotiated separately via their CP (Control Protocol) during the NCP phase (Network Control Protocol)
  • Authentication natively build in via protocols such as CHAP, PAP, MS-CHAP etc
  • Link control (~negotiable options during LCP)
  • Keepalives
  • Ability to aggregate multiple lines together into a single link such via MLP (Multilink PPP)


IP sessions don't have an authentication natively build in as there is no concept of username/password here, but with BNG deployments there are options now by having the BNG construct a username out of options of the DHCP discover or other fields such as MAC address, access interface, VLAN etc.


First Sign of Life comparison between PPPoE and IPoE sessions


Screen shot 2012-03-15 at 3.47.13 PM.png




Access Interfaces

This is the basic configuration to be setup in order to get PPPoE and IP sessions running.

Note that PPP and IPoE sessions can co-exist on the same parent interface as well as on the same subinterface.


This this is a base configuration example to setup the FSOL handling for PPPoE or IP sessions.

Screen shot 2012-03-15 at 3.48.40 PM.png


Using ambigious vlans

Originally for double tagged traffic, also known is QinQ or QiQ, we had to explicitly configure the inner and outer vlan combination for each possible termination point. In ASR9000/XR we can define ambigious ranges allowing us to specify the outer vlan only and an inner range.

The most common deployment scenario for QIQ is whereby the outer vlan represents the dslam and the inner vlan represents the subscriber, obviously configuring 64k subinterfaces is not very easy to manage and the Ambigious vlan support greatly reduces operational overhead, large configs and provides for much more simplicity:


Configuration example:


interface Bundle-Ether1.50

service-policy type control subscriber PPP_IP_PM1

pppoe enable bba-group default

encapsulation ambiguous    dot1q { any | <vlan range> }

                           dot1ad { any | <vlan range> }

                           dot1q <vlan#> second-dot1q { any | <vlan range> }

                           dot1ad <vlan#> second-dot1q { any | <vlan range> }                          


Screen shot 2012-03-15 at 4.14.34 PM.png



DHCP/IP sessions

PPP sessions have a native keepalive build in. If keepalives are not sent between the BNG and the client, the sessions are automatically torn down. IP sessions don't have a native keepalive mechanism and some implementations opted for an ICMP or ARP keepalive methodology to detect absent IP sessions as opposed to relying on (potentially long) DHCP lease timers.

ASR9000 does not have ICMP or ARP keepalives for IP sessions rather instead we have a different mechanism of lease-proxy which is elaborated on in this section.

Restart handling

Problem domain 1:

IPoE sessions are initiated upon receipt of a DHCP discover and can be terminated prior client’s IP address lease expires/is released by:

  • CoA Account-Logoff/PoD
  • Session Administratively cleared
  • Reload


Need a way to recreate session if client lease is still valid:


Screen shot 2012-03-15 at 4.06.37 PM.png


Once we have marked the Binding in the dhcp proxy on the BNG as "stale" due to one of the reasons mentioned above, a subsequent DHCP request from the client (eg on renew) is NAK'd to the client who will then faill back to a discovery mode upon which we can recreate the session.

Screen shot 2012-03-15 at 4.07.30 PM.png

To support a sort of keepalive mechanism we can shorten the lease time, which will require the session to renew its lease at half the lease time. So we effectively have a keepalive mechanism at half-lease time in this scenario.

This inherently increases the load on the dhcp server because the BNG will forward the renew requests to the dhcp server and when Acknowledged it will maintain the binding and the session.


A smoother solution is the concept of "lease proxy". This means that eventhough the server offers a lease, in this example of 40 minutes, the BNG advertises a lease to the client of a configurable time, in this example 10 minutes.

Every 5 minute interval the client will renew, but now the BNG intercepts and re-acknowledges the lease to the client, as opposed to relying on the dhcp server to ack the renew.

At half the lease time, here 20 minutes, we renew with the dhcp server to maintain proper state.


DHCP lease limits

Lease limit and Lease proxy for DHCP subscribers


Lease limit for a given proxy profile can be specified for:


per remote id:



dhcp ipv4

  profile dhcp-red proxy

  limit lease per-remote-id 1000


per circuit-id:



dhcp ipv4

profile dhcp-green proxy

  limit lease per-circuit-id 1000



per interface:



dhcp ipv4

profile dhcp-blue proxy

  limit lease per-interface 1000


Note that per circuit-id and per remote-id options are confined to any given access

interface. In other words, per circuit id limit on a given access interfaces doesn't affect

or influence the circuit-id limit configured for any other access interface


Lease proxy:

DHCP lease proxy is also known as DHCP split lease.  With this implementation, the DHCP proxy

ie., BNG router will renew the lease of the client without contacting the DHCP server. The

lease proxy value configured is assumed to be lower than the server lease. Following terminologies

are used:


i) Client<->Proxy is the client lease

ii) Proxy<->Server is the proxy lease


ASR9K is the DHCP proxy


Lease proxy benefits include:

a) shorter client lease times and longer proxy lease

b) Proxy can respond faster to renews at the network edge

c) Reduces load on centralized DHCP servers for renewal processing




dhcp ipv4

profile dhcp-red proxy

RP/0/RSP0/CPU0:BNG1(config-dhcpv4-proxy-profile)#lease proxy client-lease-time ?

  <300-4294967295>  Value of lease proxy client-time in seconds


dhcp ipv4

profile dhcp-red proxy

lease proxy client-lease-time 300


Example TAL (Transparent Autologon) use case 1

TAL or transparent autologon is one of the most commonly used access scenarios for IP sessions.

You can authenticate the user on mac address or option 82 information. The following use cases depict on how to set that up with the ASR9000 BNG implementation.

Screen shot 2012-03-15 at 4.48.06 PM.png


Example TAL use case 2

Screen shot 2012-03-15 at 4.48.30 PM.png

IP sessions and security forwarding

When your access interface is configured for IP it by nature can start forwarding IP already. A session that takes a static source ip can start forwarding traffic just fine then.

This could be a security issue and this has been done at the explicit request from some our initial adopters of A9K BNG.


Downstream traffic can only flow on AMBIGUOUS vlans when we have a session since the mapping from destination IP to mac and vlan is only held by the dhcp binding. In UNAMB scenarios, we could technically send traffic down to the sub.


UPstream traffic can be mitigated by either using:

  • uRPF
  • ACL

ACL being a far less pps implication then uRPF.

also uRPF will not necessarily capture all issues, especially when an ip source is chosen in the same range as the unnumbered or designated address of this access interface.


    When you apply an ACL, make sure it allows DHCP to go through.


    Packet trigger or unclassified Source

    This implementation allows for IPoE subscribers to be established based on

    a received packet from an IP source on the access interface when the "unclassified

    source" is enabled.  Typically, the sessions will be restricted to packets originating

    from a particular network. Multiple matching criteria may be specified to match packets

    from discontiguous networks. In addition, the packets may be subject to radius based

    AAA for a successful session bringup. An access interface may be configured to support

    both DHCP and PKT triggered sessions


    In an ideal network, the subscriber would first send an ARP request to the access

    interface and if the packet matching criteria are met, this in itself is a sufficient

    condition to bringup the IPoE session. However, if there is a burst of traffic with

    unique flows, this could overwhelm the BNG router in terms of processing each packet

    to determine if it is a IPoE (PKT) candidate. The software limits the amount of in-flight

    requests to 200. In cases where traffic rates for IPoE-PKT sessions are high (>120 pps) and

    there are also parallel DHCP based sessions creates in progress, it may be desirable to

    configure static policer on the line cards.  Based on testing results, a policer rate of

    200 per LC is shown to handle this stress condition satisfactorily.





    Wed Jun 13 14:20:40.883 PDT

    RP/0/RSP0/CPU0:BNG1(config)#lpts punt police location 0/0/cpu0  protocol unclassified rsp rate 200


    aaa attribute format ip-plus-mac

    format-string length 253 "%s:%s" addr client-mac-address




    interface Bundle-Ether10.41

    ipv4 address

    service-policy type control subscriber ipsub_policy1

    encapsulation dot1q 30

    ipsubscriber ipv4 l2-connected

      initiator dhcp

      initiator unclassified-source





    policy-map type control subscriber ipsub_policy1

    event session-start match-first

      class type control subscriber class-dhcp do-all

       1 activate dynamic-template ipsub_template

       3 authorize aaa list default identifier circuit-id password dhcp123

       5 activate dynamic-template acct-default


      class type control subscriber class-pkt do-until-failure

       1 activate dynamic-template ipsub_template

       3 authorize aaa list default format ip-plus-mac password abc123

       5 activate dynamic-template acct-default







    class-map type control subscriber match-any class-pkt

    match source-address ipv4






    type ipsubscriber ipsub_template

      ipv4 unnumbered Loopback1






    Using Control Polices

    One integral part of the BNG solution in XR is the use of control policies.

    With control policies you are able to manage the sessions life while various events on the session are triggered.

    You can handle these events or ignore them depending on your configuration and deployment needs.

    User authentication/control

    One key action in the control policy is obviously the authentication.

    These can be executed via the command under the event/class:


    • 10 authorize aaa list default identifier source-address-mac password cisco
    • 10 authenticate aaa list default


    Both of these will trigger a RADIUS access-request message, but the difference between the two is with the authorize statement we can compose the username ourselves regardless of what is received on the line, where as the authenticate statement uses the PPP chap or pap username and password received. The authenticate option for that reason only applies to PPP based sessions.


    You can define the username in the authorize statement either inline as per example above or you can construct the username via a "formatted" way:


    Screen shot 2012-03-15 at 4.02.20 PM.png


    Note that an authentication does NOT have to succeed in order for BNG to bring up the session.

    The activation of a dynamic template will create the subscriber interface regardless of the authentication result.

    A failed authentication will result in an unauthenticated state and you'll be able to apply HTTP-redirect or restricted access on the session.



    Nas identification

    Source IP

    Is by default selected of the interface's address that is used to reach the radius-server.

    This is configurable to be overriden from this default.


    RP/0/RSP0/CPU0:A9K-BNG(config)#radius source-interface <interface name>


    Or per default configuration:


    RP/0/RSP0/CPU0:A9K-BNG#show run | i radius-server

    Fri Mar 16 12:40:15.791 EDT

    Building configuration...

    radius-server host auth-port 1645 acct-port 1646

    radius-server attribute list LIST

    RP/0/RSP0/CPU0:A9K-BNG#sh route

    Routing entry for

      Known via "connected", distance 0, metric 0 (connected)

      Installed Feb 22 15:42:37.812 for 3w1d

      Routing Descriptor Blocks

        directly connected, via MgmtEth0/RSP0/CPU0/0

          Route metric is 0

      No advertising protos.


    RP/0/RSP0/CPU0:A9K-BNG#sh run int mgmtEth 0/rsP0/CPU0/0

    Fri Mar 16 12:40:26.121 EDT

    interface MgmtEth0/RSP0/CPU0/0

    ipv4 address




    Thu Mar 15 11:55:12 2012: [18848] NAS-IP-Address =



    Attribute 87 can be filled with the configuration like this:


    aaa attribute format NAS-PORT-ID

    circuit-id plus remote-id


    aaa radius attribute nas-port-id format NAS-PORT-ID



    Attribute 32 is the BNG's hostname, always and only configurable when changing the router's hostname.




    RP/0/RSP0/CPU0:A9K-BNG#sh run | i host

    hostname A9K-BNG



    Thu Mar 15 11:55:12 2012: [18848] NAS-Identifier = "A9K-BNG"



    Attribute 5


    Is computed very flexibly configured on a per session type basis:


    aaa radius attribute nas-port format e <format> [type <0-44>]


    Format (32bits): entered as a string of letters:

    Zero : 0

    One : 1

    Slot : S

    Adapter : A

    Port : P

    (Outer) VLAN Id : V

    Session-Id : U

    Inner VLAN ID: Q





    ETHERNET   15

    PPPOEOE   32







    If type is omitted it will apply and be used for any session without a more specific type definition.


    Control policy Events, Classes and Actions


    The folllowing pictures shows how everything ties together in a control policy:

    Screen shot 2012-03-15 at 4.17.20 PM.png


    The diagram below shows where the various events would be triggered for a PPPoE session.

    Note that the session activate event is only applicable to PPP sessions.


    You need to make sure that the session-start event has a template defined with the lcp paramters

    which are used during LCP.


    Screen shot 2012-03-16 at 11.08.40 AM.png


    the power of the solution, amongst many others, is the differentiation you can do between authentication failures as well as no response,

    so you can act upon a faulty username differently then a radius-server not responding.


    Failover for that reason can be embedded in the control policy like this:


    event session-activate match-first

      class type control subscriber CLASS do-until-failure

       10 activate dynamic-template TPL

       20 authenticate aaa list default


    event authentication-failure

       class ...

       10 "apply http redirect"


    On authentication failure we an apply a layer 4 redirect service while keeping the session active.


    event authentication-no-response

       10 authenticate aaa list failover-list


    If there was no response from radius, we can try a different radius-server list


    Using Class-Maps

    The class map definition allows you to control how the event triggered is handled.

    Either the event is handled for the first class that is matched, or ALL classes for this event are evaluated as part of the event definition directive.



    class-map type control subscriber match-any IP_SUB

    match protocol dhcpv4

    ! The above would match specifically on IP subscribers only


    class-map type control subscriber match-any PPP_SUB

    match protocol ppp

    ! This example would match specifically on PPP subscribers only


    Using these example classes allows you to have a single control policy that can handle events and use different actions per access category.


    Available Match Criteria:

    • Domain name: domain <string>
    • Protocol: protocol { dhcpv4 | ppp }
    • Source address: source-address { ipv4 | mac }
    • User name: username <string>
    • Authentication Status: authen-status { authenticated | unauthenticated }
    • To negate match criteria: not <>


    Available Match policies (as part of the class-map definition):

    •       match-any: match any of match clauses
    •       match-all: match all match clauses


    If you only have 1 match clause in your class-map it obviously doesn't make a difference whether you choose match-all or match-any.


    Also the class-maps allow for very extensive control of the event handling whereby you can handle a particular event differently for an unauthenticated ppp subscriber vs an authenticated ip subscriber or any combination of that of course!


    Handling failed authentications

    When an authentication fails for a subscriber against radius, we can still bring the user up by activating a dynamic template.

    So the order of actions executed during an event is very important along with using do-until-success/failure etc.


    For instance the following actions:

    event session-start do-all

    10 activate dynamic-template TPL

    20 authorize aaa list default mac-address password cisco

    would bring up the subscriber even with failed authentication.


    In this example where the actions are effectively reversed:

    event session-start do-until-failure

    10 authorize aaa list default mac-address password cisco

    20 activate dynamic-template TPL

    would not bring up a user after authentication failure.


    Alternatively, you can pull in the event for authorization failure and disconnect the service like this:

    event authorization-failure do-all

    10 disconnect


    Or you can use the authorization failure to apply HTTP-Redirect and start a timer, so effectively allowing the user to login within that time before he is getting disconnected.


    Account Logon

    If the user failed authentication and has a restricted access service applied, we can force the user to go the web portal to provide credentials and try to login again, pay their bill etc.


    The model is here that the user goes to a web page to provide credentials that are then send via a coa account logon to the BNG.

    The BNG will generate an access request to authenticate using these credentials.


    If it succeeds new attributes can be send in that access-accept to remove the restricted access or HTTP-R service.

    Screen shot 2012-03-16 at 1.22.28 PM.png



    How IP sessions /DHCP interact with AAA

    Screen shot 2012-03-15 at 4.49.09 PM.png






    Address assignment options

    Obviously in order for a subscriber to have access to the network, an address has to be handed out. There are different options available for this that'll be listed out here.

    Local pools (PPPoE)

    When you have PPPoE based sessions, the easiest implementation is to define a local pool from which the addresses are handed out to the subscriber during IPCP


    The associated configuration is:

    pool vrf default ipv4 POOL



    Multiple ranges can be provided and addresses in that range can be excluded.


    Which can be monitored via this command:

    RP/0/RSP0/CPU0:A9K-BNG#show pool ipv4


                  Allocation Summary


    Used: 1

    Excl: 0

    Free: 65278

    Total: 65279

    Utilization: 0%


      Pool         VRF      Used   Excl   Free   Total

    ----------  ----------  -----  -----  -----  -----

          POOL     default      1      0  65278  65279


    The Pool can either be referenced directly on the dynamic template which is activated to the subscriber during its event handling in the control policy like this:


    This is the template that holds the base configuration for subscribers when this template gets activated on the session:


    type ppp TPL

      ppp authentication chap

      ppp ipcp dns

      ppp ipcp peer-address pool POOL

      ipv4 unnumbered Loopback1000


    This template can then be referenced in an event handling of the control policy as with this example:


    policy-map type control subscriber sub

    event session-start match-first

      class type control subscriber CLASS do-until-failure

       10 activate dynamic-template TPL


    Alternatively the POOL can also be referenced via Radius Attributes during the Access-Accept as per following example:


    • With a Cisco-Avpair        Password = "cisco"

          Service-Type = Framed-User,

          Framed-Protocol = PPP,

          Cisco-avpair = "ipv4:addr-pool=POOL",


    • Using an Ascend Attribute (number 218):
         Ascend-Assign-IP-Pool = POOL


    • Using an IETF Attribute (number 88):

              Framed-Pool = POOL


    • Alternatively, but uncontrolled is the use of the Framed-IP-Address magic number which will instruct the NAS to do a "pool pick" from any pool available.


    Using this method of locally defined pools on the BNG is by far preferred because it allows us to create a summary route and advertise the pool in its whole. This reduces significant amount of routing updates, but has the limitation that a full block is assigned to the BNG regardless of whether it needs it or not.


    Pool advertisement

    Can be done via the following methodology:

    First define a summary route for the pool range:


    router static

    address-family ipv4 unicast Null0


    Next inject that summary route into your eg IGP via a redistribution command

    router ospf CORE

    redistribute static


    When users come online they will have a /32 in the routing table which is then followed for forwarding rather then the summary route to NULL0.

    Another "disadvantage" of this methodology is that you'll be drawing all 199.1.x.x traffic to the BNG regardless of whether there is a session or not. However the ingress LC's NPU will drop the packets in hardware in that case.


    Radius based pools (PPPoE)

    Instead of assigning a pool on the BNG, you can also outsource the pool management to a radius-server.

    IT is recommended to have the radius server select a pool per BNG device, this in order to keep the model of summary advertisement.

    If the pool attributes are distributed between different BNG's, you're required to inject the /32's which will put unnecessary burden on your IGP.

    In this case you probably want to consider STUB areas to keep the /32's only floating in your OSPF STUB area and summarize them at the area border.


    Radius based pools rely on the accounting mechanism from AAA to learn whether addresses are in use or not.

    This requires a strong Accounting back end on the radius-server and obviously proper delivery of your AAA records.


    Don't implement radius pools without AAA System accounting.


    Static address assignment (PPPoE/IPoE)

    Requires you most of the times to inject /32's into the IGP for proper routing if not part of a summary.

    The way to achieve this is via the radius attribute Framed-IP-Address (IETF number 8).


    IP sessions

    IP sessions address assignment is done via a DHCP Server. IOS-XR release 4.3.0 will come with an on board dhcp server.

    In today's model the dhcp server is responsible for handing out addresses which are picked based on the "giAddr" field in the dhcp discover which is filed in by the DHCP Proxy component of the ASR9000.

    Static addresses are defined within the dhcp server.


    The advertisement is here more simple since the unnumbered interface to the subscriber session provides the subnet it is serving hence the inclusion of this unnumbered interface in your IGP will take care of the proper routing:


    In this example we set the dhcp server to, the giAddr for pool selection is set to the red value to instruct the dhcp server where to pick an address from. The giAddr selection in this example is based on the dhcp option 60, vendor Class which is matching a hex string.


    dhcp ipv4

    profile AutoSelectGiaddr proxy

      class MATCHALL

       match option 60 hex 68656C6C6F mask 0

       helper-address vrf default giaddr


      class HardPhone1

       match option 60 hex 4861726450686F6E6531 mask 0

       helper-address vrf default giaddr


      class HardPhone2

       match option 60 hex 4861726450686F6E6532 mask 0

       helper-address vrf default giaddr


      relay information option

      relay information policy replace

      relay information option remote-id testme

      relay information option allow-untrusted


    interface Bundle-Ether100.2 proxy profile AutoSelectGiaddr


    This combined with the following template configuration, the subscriber is unnumbered to loopback interface number 12.



    type ipsubscriber IPSUB

      ipv4 unnumbered Loopback12


    Which has then address in the same pool range as the giAddr and the dhcp server will want to set the default-router option to this value.


    RP/0/RSP0/CPU0:A9K-BNG#sh run int lo12

    Fri Mar 16 10:52:33.265 EDT

    interface Loopback12

    ipv4 address


    Including Loopback 12 (passively) in your IGP will provide for proper downstream routing!


    More information on the DHCP configlet is in the BNG training guide


    Applying configuration to a subscriber

    Configuration can be applied to a subscriber via 3 methods and in order of preference from lowest to highest:

    1. Dynamic Template
    2. Radius/Access-Accept (also known as Policy PULL)
    3. COA (also known as Policy PUSH)


    So that means that local template configuration can be overridden by RADIUS, which can be overidden by COA.


    Here are a few examples on how to match a CLI Configuration to a radius attribute, full documentation is available in the XR configuration guide.


    Screen shot 2012-03-15 at 4.44.23 PM.png

    Screen shot 2012-03-15 at 4.44.35 PM.png




    PPP timer configuration

    It is very important for any BNG implementation to properly define your PPP timer configuration to make sure that sessions establish, also under load and that we are not prematurely timing out, or working on establishment while the peer already gave up.


    The picture below shows the different stages of PPP and which timers apply that need evaluation.


    Screen shot 2012-03-16 at 11.10.31 AM.png


    The referenced timers can be configured on the dynamic template for PPP subscribers:


    dynamic-template type ppp <tmpl_name>

    ppp max-configure <#>

         ppp max-failure <#>

         ppp timeout retry <sec>


         ppp lcp delay <sec>

         ppp lcp renegotiation ignore

         ppp authentication { pap | chap | ms-chap }

         keepalives { disable | <sec> }


         ppp max-bad-auth <#>

         ppp timeout authentication <sec>


         ppp ipcp peer-address pool <poolname>

         ppp ipcp mask <mask> 

         ppp ipcp { dns | wins } <server ip>

         ppp ipcp renegotiation ignore


         ipv4 unnumbered <interface>



    Quality of Service

    QOS application to the session can be done via static configuration in the dynamic template, or the policy-map name can be referenced via Cisco AVP's in RADIUS access-accept or COA requests.

    This requires the configuration of the policy-map to be present in the XR configuration.


    QOS can be applied at the port level, vlan level (subinterface)  and at the session level with classes in a hierarchical manner.

    This constitutes the 4 layers of QOS.


    The following pictures shows the basics of the 4 layer QOS:


    Screen shot 2012-03-16 at 11.03.07 AM.png

    Which can be further expanded on in this picture:


    Screen shot 2012-03-16 at 11.03.17 AM.png



    Parameterized QOS

    Parameterized QOS is a very powerful option in ASR9000 BNG. It allows you to construct the policy-map and its values from the AAA server.

    pQOS only requires the class-map definitions to be present in XR configuration.


    You can setup hierarchical policy-maps without a problem using pQOS.


    Note however that if you have defined a static policy-map via configuration to the dynamic template or from radius

    then you cannot override it or modify it with pQOS.

    If you desire to use pQOS the initial policy-map needs to be pQOS'd also.


    You can modify pQOS policies on a per class basis and while the session is active add or remove classes dynamically as you go!!

    To see pQOS in action and the benefits see this video on demand BNG demo on YOUtube


    VSA definition

    Understanding the format of the Vendor Specific Attribute for Parameterized QOS (pQOS).

    VSA(9-1)”qos-policy-in:add-class( <target-specifier>,(<class-list>),<qos-action-params> )”


                             VSA 9,1 is Cisco-Avpair


    sub – The QoS policy attached to the subscriber session. This implies that the CoA/Access-Accept target must be a subscriber session.



    This example identifies the class “class-default” on the parent-policy.


    This example identifies a leaf class “voip”. This class will be added to or removed from a nested child policy specified under the class “class-default” of the policy attached to the target.


    This example specifies a leaf class “voip-1”. This class will be added to or removed from a nested child policy specified under the class “voip-aggregate” of the policy which is in turn nested under “class-default” of the policy attached to the target




                   See this next section on how to map IOS-XR MQC (modular Qos configuration) actions to the parameterized QOS equivalent.


    Supported pQOS actions:


    Screen shot 2012-03-16 at 10.09.19 AM.png



    CLI Equivalent: police <bps> <burst-normal> <burst-max> <burst-size> conform-action <action> exceed-action <action> violate-action <action>


    VSA value: qos-policy-in:add-class(sub,(voip),police(200000,9216,0,0,transmit,drop,drop) )



    CLI Equivalent: shape <shape-rate>

    VSA value: qos-policy-out:add-class(sub,(class-default),shape(14700))




    For complete COA examples check the Change of Authorization document



    Multicast and video distribution


    Coming soon!!



    Call Admission Control

    Unlike IOS, ASR9000/XR BNG has no explicit call admission control configuration as that is natively build into the system.

    When resources are running slow or the system experiences internal back pressure to do slow responses on function calls

    the number of in flight sessions increases which are then throttled back down at the access interface.


    Most of the time you may want to control the number of in flight sessions so you can streamline the number of radius access requests that are being sent to the RADIUS server.


    The way to monitor and control the in flight sessions is via this command :


    RP/0/RSP0/CPU0:A9K-BNG#show pppoe summary total location 0/rsP0/CPU0

    ....<output omitted>...


    Flow Control


    Limit                    2000

    In Flight                   0

    Dropped                     0

    Disconnected                1

    Successful                  9


    Limit means the number of in lfight sessions you want to control on a per node basis. A node constitutes a pppoe processing entity which is either the LC for phyiscal interface based sessions or the RSP when using bundle interfaces. This number is configurable via the followingcommand:

    RP/0/RSP0/CPU0:A9K-BNG(config)#pppoe in-flight-window 2000

    Default value is 200, recommended for RP-based subscribers. Recommended value for LC-based subscribers is 50.


    In Flight is the number of sessions we are currently handling and have not fully established yet. A fully established session is the signal from the session control entity that the subscriber interface is up and forwarding


    Dropped are the sessions when the in flight session number exceeds the Limit set.


    Disconnected how many sessions ahve been disconnected for normal reasons, eg send a PADT etc.


    Successful is the number of sessions that we successfully established over time on this node





    This feature supports throttling of access (authentication and authorization) and accounting records

    that are sent to the radius server. Throttling rate can be configured separately for access

    and accounting requests. When the threshold is reached for a server, no more requests of that type

    will be sent. A retransmit timer is started when the threshold limit is reached. After expiry of

    the retransmit timer, the queue is checked to see if the outstanding requests is less than the

    configured limit. If so, then the request is sent out to the radius server


    AAA throttling can be configured globally or at the server group level. Throttling configured

    for the server group will take precedence




    radius-server throttle access 100 access-timeout 3 accounting 150

    aaa group server radius my-grp

    server auth-port 1812 acct-port 1813

    server auth-port 1812 acct-port 1813

    throttle access 200 access-timeout 3 accounting 120




    Subscriber Services

    Services constitute a set of features under a common umbrella.

    These features are enabled together constituting the service defintion.


    For instance you can allow users to access or deny parts of the network, or modify its qos parameters.


    Services are defined via the dynamic template in IOS-XR.



    dynamic-template type { ppp | ipsubscriber | service } <tmpl_name>



    There are 3 types of templates:

    ppp: for configuration on PPP sessions (both PTA and LAC)

    ipsubscriber: for configuration on IpoE sessions

    service: contains configuration commands for all types of sessions

    Dynamic templates allow for inline modifications, changes take effect immediately on all sessions using template, with the exception: unmutable config options (e.g session IP address)


    The following is an example of a service definition:



    type service SERVICE_1

      service-policy output testme

      ipv4 access-group lab-video ingress


    The following set describes a few RADIUS/COA templates to activate or deactivate services on a subscriber.




    RADIUS attributes set to achieve that operation

    Account Logon

    attribute   44 “<string>”

    attribute 1 "<username>

    Cisco-avpair = "subscriber:password=<subscriber password>

    Cisco-avpair = "subscriber:command=account-logon"

    Account Logoff

    attribute   44 “<string>”

    Cisco-avpair = "subscriber:command=account-logoff"

    Account Update

    attribute   44 “<string>”

    Cisco-avpair = "subscriber:command=account-update”

    <radius   attributes to set/update>

    Service Activate

    attribute   44 “<string>”

    Cisco-avpair = "subscriber:command=activate-service"

    Cisco-avpair = "subscriber:service-name=<service-name>”

    Service De-activate

    attribute   44 “<string>”

    Cisco-avpair =  "subscriber:command=deactivate-service"

    Cisco-avpair = "subscriber:service-name=<service-name>”


    Attribute 44, or accounting session ID is always used for fastest lookup of the subscriber session.

    The subscriber password is a special encoded string by providing a seed authenticator together with a hashed MD5 password of that seed.

    COA tool is available on the forum for Windows, MAC/OSX and Linux.


    Redundancy and fan out

    Redundancy is obviously very important these days and while IOS-XR supports process restart and process crashes without affecting the subscribers there are additional options available to increase your density and protect against failures.


    Cluster and Satellite are part of the ASR9000's nV concept (Network Virtualization).


    Cluster is the concept of binding two physical chassis together into 1 logical unit. The control plane is extended via RSP on board 1 or 10G interfaces while the data plane is extended via 1 or more physical interfaces on the linecards.


    So when building a bundle from the access side, if you link them to each individual chassis you'll have an active active bundle with failover between chassis!


    Also the physical ASR's don't need to be on the same location. Only the control plane extension needs to have minimum latency (~<20msec).


    Screen shot 2012-03-16 at 11.20.05 AM.png


    Screen shot 2012-03-16 at 11.53.40 AM.png



    Allows for port extension into a simple 1RU chassis with large 1G port fan out.

    The Satellite connects via 1 or multiple 10G uplinks ot the ASR9000 host.

    You can statically pin ports to an uplink or share an uplink via a bundle to multiple ports.

    Satellite interfaces appear in teh ASR9000 config as if they were physically on the ASR9000.



    Satellite interconnection options:

    Screen shot 2012-03-16 at 11.56.32 AM.png


    Configuration example:



    satellite 100 •ß define satellite ID

    description my lovely satellite

    type asr9000v


    satellite 101 •ß define satellite

    description your lovely satellite

    type asr9000v


    interface TenGigE 0/2/0/2


       satellite-fabric-link satellite 100


           GigabitEthernet 0/0/0-9



    interface bundle-ethernet 10


       satellite-fabric-link satellite 101


           GigabitEthernet 0/0/10-19


    On top of these bundle interfaces you can enable BNG.



    Cluster and Satellite together


    Screen shot 2012-03-16 at 11.58.21 AM.png


    Wholesale Models

    Most of the documentation here has been talking about locally terminating the subscribers on the BNG for regular access. Obviously there deployment models whereby a provider may like to just provide the initial termination on behalf of the wholesale provider.


    Per access technology there are different options available.


    Screen shot 2012-03-16 at 12.01.56 PM.png



    PPP sessions can either be locally terminated, also known as PTA, or forwarded to the wholesale provider with 2 main options:

    • L2TP tunneling
    • RAMPLS


    IP sessions

    IP sessions only have one option for wholesaling which is inserting the subscribers in a VRF and using MPLS VPN to transport the data traffic to the wholesale provider.


    (Remote Access into MPLS-VPN) Is the concept of terminating the subscriber sesisions locally on the BNG and insert them in a specific vrf. This vrf is a separate routing context and using MPLS-VPN to transport the users traffic to the wholesale provider.


    Layer 2 Tunnelling Protocol is the concept of transporting the PPP session over to the wholesale provider's LNS.

    ASR9000 can only function as LAC which basically means that after authentication we are creating a tunnel or inserting the user into an existing tunnel over to the LNS.


    L2TP has as key advantage that the subscriber's PPP session is sent over to the LNS. This allows the LNS to have full control over the PPP session including authentication.

    RAMPLS doesn't have such an option as the only Authentication stage is done on the BNG one time.

    The concept of "double dip" is a phenomenal extension allowing you to use a local radius server on teh BNG and then contact the wholesale provider's AAA server and merge the profiles together:


    Screen shot 2012-03-16 at 12.12.26 PM.png


    On the ASR9000 BNG you can filter the attributes from the retailer to make sure that they don't override the user's vrf for instance like this:


    radius-server attribute list RETAILER_X_ATTR_LIST

    attribute <accepted or rejected attribute-list>


    aaa group server radius RETAILER_X_SG

    authorization reply { accept | reject } RETAILER_X_ATTR_LIST


    server-private auth-port 1645 acct-port 1646




    This server group referenced can be used in a control policy as an additional authenticate/authorize action under the session-start event for instance.



    High Scale Deployment Best Practices

    If your BNG deployments is high scale, these are some of the best practices to follow:

    ICMP Unreachables

    ICMP unreachables can significantly increase the CPU utilisation on line cards, especially if security access lists are enabled on seubscriber interfaces. You can disable the generation of ICMP unreachables in subscriber access interface configuration or dynamic template configuration:

    ipv4 unreachables disable

    or in radius profile:



    In deployment scenarios with IP unnumbered and loopback interfaces with multiple secondary IP addresses, the ARP table size grows very quickly. Since on subscriber interfaces the MAC to IPv4 mapping is known through either DHCP or PPPoE, ARP entries are not required. On XR release 5.3.3 and later you can prevent the creation of ARP entries associated with subscriber interfaces by configuring:

    subscriber arp scale-mode-enable

    For more information on high scale ARP, refer to:

    Related Information


    Special thanks to Sabrina Pittarel for some of the great visualizations seen in this document

    and Nanjangud Sreekanta Prasad providing packet trigger and dhcp lease limit sections




    Xander Thuijs, CCIE #6775

    Sr. Tech Lead ASR9000

    Cisco Employee
    Cisco Employee

    Hi Vladimir,

    we dont have time based ACL in xR today, but there are some options to achieve this!

    option 1) using COA to modify the applied policy to a session.

    option 2) using a script to modify the classification ACL that is used in the pmap that uses a particular shaper/policer. the acl modification basically leads to the classification of traffic in a different class with that different bw/policer.

    Both options have pro's and con's:

    1) with 30-50 requests per second on COA it can take a few seconds if the number of sessions that need to be modified.

    2) we know of some scale issues that we're fixing whereby the acl reconfig used in a class-map could result in a timeout or error induced by the QOS management agent (MA, part that takes the update/call from a "source" eg cli, xml etc) or the Execution agent (EA, the part that takes the calls from the MA and programs the hardware with the updates).

    True timebased ACL we're not tracking as an official feature yet.

    Another option instead of using dynamics like this we can also leverage a portal approach whereby a user wanting some turbospeeds can request that via the portal which induces the COA to activate that service, we can start a timer on that service to automatically deactivate that after a certain time.

    Advantage of this approach is that it is user driven and reduces the COA overhead of modifying sesions that may not need/want this and it improve the overall user experience because the extra availalble BW is not given/shared to others that dont have that service activated.



    Hello Xander,

    Thanks for your answer!

    We need to modify policies of 15-20K subscribers simultaneously at midnight and move it back at 8.00 a.m.. ASR1000 makes it easy with time-based ACL. I think for ASR9001 with COA it will be hard work. 

    Do you have any plans in road map to include time-based ACL in IOS XR code? If so, in which version it will be included?

    Thank you!

    Hi Xander,


    Have a question on how to clear or delete RADIUS profile templates cached on BNG?

    Following is bothering me, after RADIUS server returns Service_Name to apply to subscriber in Access-Accept message, BNG asks in another Access-Request for Service definition (AVP's). RADIUS server responds with Service AVP's (inacl, ....) and BNG caches it locally. This is giving me trouble since I'm in testing phase and I often change Service definition.

    This BNG behavior make sense since Service definition would not change often or not at all. But in testing phase it would be great to have a way to clear BNG cache.


    Appreciate any suggestion.





    Cisco Employee
    Cisco Employee

    good question! not sure! reload doesn't count right? :) I'll check and get back.


    Frequent changes will took ages to test it with reload.

    Any other way would be most helpful.





    Thulasirajan Alagarsamy
    Cisco Employee
    Cisco Employee



    There is no way to clear the service cache using command and there is an open enhancement. However we can update the cached service definition with the help of coa service-update. And all the users which are using the service will get updated values.


    echo "cisco-avpair+='subscriber:command=service-update',Cisco-avpair+='subscriber:service-name=pQOS_SVC_1MIN',Cisco-AVPair+='Method-List=default',Cisco-AVPair+='ip:qos-policy-in=add-class(sub,(class-default),police(1085))',Cisco-AVPair+='ip:qos-policy-in=add-class(sub,(class-default,BROADBAND_VOZ),police(512,transmit,drop),set-mpls-exp-imposition(4))',Cisco-AVPair+='ip:qos-policy-in=add-class(sub,(class-default,BROADBAND_CRITICOS),set-mpls-exp-imposition(2))',Cisco-AVPair+='ip:qos-policy-in=add-class(sub,(class-default,BROADBAND_BUSINESS),set-mpls-exp-imposition(2))',Cisco-AVPair+='ip:qos-policy-in=add-class(sub,(class-default,class-default),set-mpls-exp-imposition(0),set-ip-dscp(0))'" | /usr/local/bin/radclient -x coa cisco

    Sending CoA-Request of id 173 to port 1500

            Cisco-AVPair += "subscriber:command=service-update"

            Cisco-AVPair += "subscriber:service-name=pQOS_SVC_1MIN"

            Cisco-AVPair += "Method-List=default"

            Cisco-AVPair += "ip:qos-policy-in=add-class(sub,(class-default),police(1085))"

            Cisco-AVPair += "ip:qos-policy-in=add-class(sub,(class-default,BROADBAND_VOZ),police(512,transmit,drop),set-mpls-exp-imposition(5))"

            Cisco-AVPair += "ip:qos-policy-in=add-class(sub,(class-default,BROADBAND_CRITICOS),set-mpls-exp-imposition(1))"

            Cisco-AVPair += "ip:qos-policy-in=add-class(sub,(class-default,BROADBAND_BUSINESS),set-mpls-exp-imposition(1))"

            Cisco-AVPair += "ip:qos-policy-in=add-class(sub,(class-default,class-default),set-mpls-exp-imposition(0),set-ip-dscp(1))"

    rad_recv: CoA-ACK packet from host port 1500, id=173, length=20


    Only limitation is we have to use entire content incase of pqos services with new value otherwise it will replace it with latest attribute/value.





    Alireza Karimi


    First thank you so much for this article

    Could you please tell me what is the minimum requirements (hardware + software + license(s)) to be able to have 128000 PPPOE sessions on ASR9000?

    Best regards

    Cisco Employee
    Cisco Employee

    hi! and thanks for the nice comment!

    for 128k sessions you need:


    2xMOD80-SE with any MPA of liking (2x10G, 4x10G, 20x1G, 1x40G)

       or higher density cards such as mod160, 24x10 or 36x10).

    software recommended is XR 5.2.4

    this assumes linecard terminated subscribers (so not on bundles).

    64k subs per LC (hence 2 minimum) and 32k per NPU (mod80 has 2 NPU's).

    you can use interfaces from the MPA for core access (unless using the 1x40G since all interfaces in that case would be terminating subscribers)

    Since you have 2 LC's, you need a 9904 minimally. For growth: 9006 or 9010 chassis

    This is all bare minimum, would recommend to go with 3-4 LC's (budget depending of course) and 9006 chassis.

    In terms of lics, any combination that totals up to be 128k .




    Alireza Karimi

    Thank you so much Alexander

    Do I have to use two PA like 2x10G on each MOD80-SE to reach 64K on each line card? or is it possible to use just one PA on each MOD80-SE and leave the other one empty? (budget problem)

    I mean does the number of users also depend on number of PAs or is it just related to the number of Line Cards?

    And about the license, do I have to buy 16 of A9K-BNG-LIC-8K to reach 128k?


    Best Regards


    Cisco Employee
    Cisco Employee

    yeah you need 2 MPA's since on the mod80, you have 2 NPU's. each npu serves a bay.

    since an npu can only hold 32k, you need both of them in use to reach 64k per LC.

    you could also consider/price out a mod160 (2 npu's per bay) and a single MPA. but I dont know how that price wise stacks up, eg mod160+1 MPA ?= MOD80+2 MPA.

    the number of users is a factor of LC's (64k tops) and NPU (32k per NPU tops).

    as for lic, yeah you'd take a number of the 8k lics to get to 128k.

    but check with your account reps also to see if there are bundles or anything available, but since I dont deal with shift-4's (=$) I need to defer this to those folks in the know (they are not on the forums).



    Alireza Karimi

    Thanks Alexander

    It was really really helpful

    And one more question:

    Can I have per session queuing in LC based subscribers (for example via policy maps)?

    Best Regards

    Cisco Employee
    Cisco Employee

    Sure thing! totally possible, you can have ACL, QOS no problem on LC based subs also.



    Hey Xander, how are you? Hope you are doing good!

    We thought everything was clear regarding the sessions in ASR9K BNG but your last commend confused us :-s

    You said that "the number of users is a factor of LC's (64k tops) and NPU (32k per NPU tops)"

    In you have answered in Tasso's question that we can reach 128K in a single MOD160.

    Could you please clarify which is true?

    Also, we are running 5.1.3 in our BNGs as per an older recommendation. I guess you suggest us to move to 5.2.4. Any things to have in mind when upgrading?

    Thanks and Regards,


    Cisco Employee
    Cisco Employee

    hi dimitris,

    ah yeah I remember our discussion and I can see where the confusion comes from, let me clear that up:

    At this moment the max sessions per LC is 64k, regardless of LC type.

    Theoretically the LC's with more then 2 NPU's (that is anything but the asr9001 or MOD80) can go beyond that number, but that is not something we have made to work yet. this because of testing limitations and mem optimizations since the LC CPU memory is not large enough to go much beyond the 64k.

    I think one of the responses in the other discussion: "if you have 4 MOD160's then you can get to 512k! But that number is in the works for testing etc. The HW will support it."

    I guess I should have added the word "theoretically" there, but I thought from the comment on testing in that second line made it clear that today we can't do it yet, but the hw has the capability...


    Thanks for the quick answer and the clarifications!

    Any suggestions regarding the upgrade from 5.1.3 to 5.2.4?

    Getting Started

    Find answers to your questions by entering keywords or phrases in the Search bar above. New here? Use these resources to familiarize yourself with the community:

    Recognize Your Peers
    Quick Links