Showing results for 
Search instead for 
Did you mean: 

ASR9000/XR Netflow Architecture and overview




This document decribes the ASR9000 netflow architecture.

It provides a basic configuration how to set up netflow and what the parameters are for scale and how netflow is implemented in the ASR9000/XR

Basic configuration for netflow

The basic configuration for netflow consists of an

Flow monitor map

An exporter map

and a sampler map.


The Flow monitor MAP pulls in the Exporter map


On the interface you want to enable netflow on, you pull in the monitor map and the sampler map.


flow monitor-map FM
record ipv4
exporter FE
cache permanent
cache entries 10000

! cache timeouts define how frequently we export what, max of 1M per LC

cache timeout active 2
cache timeout inactive 2


flow exporter-map FE
version v9
  options interface-table timeout 120

! these 2 define the exports of the sample map and interface table to theflow collector for sync'ing indexes

! to names etc.
  options sampler-table timeout 120

transport udp 1963


source <interfacename>


sampler-map FS
random 1 out-of 1


interface GigabitEthernet0/0/0/20
description Test PW to Adtech G4
ipv4 address

flow ipv4 monitor FM sampler FS ingress



Scale parameters for netflow


Ø    Trident: 100kpps/LC (total, that is in+out combined)
       Typhoon: 200kpps/LC 
Ø   1M records per LC  (default cache size is 64k)
Ø   50K Flows per sec export per LC
Ø   Sample intervals from 1:1 to 1:64k
Ø   Up to 8 exporters per map, vrf aware



Netflow support




ØIPv4 (unicast/multicast)


ØIPv6 (unicast/multicast)
ØMPLS (with or without IPv4/IPv6 fields)



Netflow is not hardware accelerated in the ASR9000 or XR for that matter, but it is distributed.

What that means is that each linecard individually runs netflow by itself.

Resources are shared between the interfaces and NPU's on the linecard.




When you have 1 interface to one NPU on one linecard enabled for netflow, the full rate is available to that interface, which is 100k pps for trident and 200k for typhoon.

When you enable 2 interfaces on the same NPU on the same LC, then both interfaces share the 100k pps (trident) or 200k pps (typhoon)


When you enable 2 interfaces on 2 different NPU's, then both NPU's share the total rate of 100k/200k amongst them giving each NPU 50k or 100k depending on the LC type.


Packet flow for netflow

•Once they pass through the sampling policer, the ucode extracts data from the header fields and sends to LC CPU to construct a flow record.
•The LC CPU sends the flow record to netflow cache on the LC.
•The flow records remain in the LC cache untill they are aged due to either timer expiry or cache exhaustion.
•There are two timers running for flow aging, the active timer and the inactive timer.

Inside the LC CPU

Screen Shot 2012-09-26 at 4.29.52 PM.png



Netflow Cache size, maintenance and memory

In IOS-XR platforms, it is the LC processor memory that holds the netflow cache.

NetFlow Cache is a Section of memory that stores flow entries before they are exported to external collector.

The ‘nfsvr’ process running on the linecard, manages the netflow cache.


Memory usage

The memory used can be monitored via this command:

           show flow monitor FM cache internal location 0/0/CPU0


          Memory used:                             8127060



Total memory used can be verified by checking the process memory util of "NFSVR"


show processes memory location 0/0/CPU0 | inc nfsvr

257       139264     65536     73728   12812288 nfsvr


Sample usage

The memory used with the cache size of default 64k entries for ipv4 & MPLS is about 8MB & for ipv6 is about 11MB.

The memory used with the cache size of maximum 1M entries for ipv4 & MPLS is about 116 MB & for ipv6 is about 150MB.

The memory used with cache size of maximum 1M entries (default is 65535) is about 116 MB per ipv4 flow monitor .

If ‘n’ ipv4 flow monitors are used all with maximum 1M entries, the memory used would be n x 116 MB.


Cache Size

The default size of the netflow cache is 64k entries. The maximum configurable size of the netflow cache is 1M entries.


Configuration to set the cache entries to ten thousand looks as follows:


flow monitor-map FM

cache entries 10000



95% of configured cache size is the high watermark threshold. Once this threshold is reached, certain flows (longest idle ones etc) are aggressively

timed out. XR 4.1.1 attempts to expire 15% of the flows.


The show flow monitor FM cache internal location 0/0/cpu0    command will give you the data on that:      


Cache summary for Flow Monitor :

Cache size:                         65535

Current entries:                       17

High Watermark:                     62258


this syslog message means that we wanted to add more entries to the cache than what it could hold. There are a few different reasons and remediations for it:

- the cache size is too small, and by enlarging it we can hold more entries

- the inactive timeouts are too long, that is we hold entries too long in the cache not getting aged fast enough

- we have the right size cache, and we do export them adequately, but we are not getting the records out fast enough due to volume, in that case we can tune the rate limit of cache expiration entries via:


flow monitor <name> cache timeout rate-limit <time>


Permanent cache


The permanent cache is very different from a normal cache and will be useful for accounting or security monitoring. The permanent cache will be a fixed size chosen by the user. After the permanent cache is full all new flows will be dropped but all flows in the cache will be continuously updated over time (i.e similar to interface counters).


Note that the permanent cache uses a different template when it comes to the bytes and packets.

When using this perm cache, we do not report fields 1 and 2, but instead use 85 and 86.

Fields 1 and 2 are “deltas” 85 and 86 are "running counters".

In your collector you need to "teach" it that 1 and 85, 2 and 86 are equivalent.




Number of flows : Total number of unique flows going through the interface in a given time period.
Cache timeout values : In general, the longer the timers, the larger the needed cache size. Short timers dictate that most records will be removed due to aging.
Average flow duration : The longer the average flow duration, the longer the timers that are needed, and thus the larger the cache.
Sampling Rate : Lower the sampling rate (i.e. lower the X, for 1:X), more flows would be populated in the cache and hence larger cache size is needed.

Which packets are netflowed?


All packets subject to sampling, regardless or whethe they are forwarded or not are subject to netflow.

This includes packets dropped by ACL or QOS policing for instance!


A drop reason is reported to NF..
* ACL deny 
* unroutable 
* policer drop 
* WRED drop 
* Bad IP header checksum
* TTL exceeded 
* Bad total length 
* uRPF drop




IPV4SrcAddr      IPV4DstAddr      L4SrcPort  L4DestPort IPV4Prot IPV4TOS InputInterface  ForwardStatus        ByteCount    PacketCount  Dir      3357       3357       udp      0     Gi0/1/0/39      DropACLDeny          415396224    8654088      Ing




As described in the architecture section, the total sampling capability depends on the number of interfaces having netflow enabled.

It shaped up to be something like this table:

# of NPs Enabled for Netflow

Policing Rate Per Trident NP (Unidirectional)

Policing Rate Per Typhoon NP   (Unidirectional)














Punt Policing of netflow packets


All packets that exceed this rate are dropped by the punt policer.

You can verify that by the controllers np counters command.


show controllers np counters all

                Node: 0/0/CPU0:



Show global stats counters for NP0, revision v2


Read 67 non-zero NP counters:

Offset  Counter                                         FrameValue   Rate (pps)



934  PUNT_NETFLOW                                     18089731973        6247

935  PUNT_NETFLOW_EXCD                                       6245           0


The _EXCD depicts that the police rate had been exceeded.

This means that you likely have to increase your sampling interval.



Monitoring netflow



sh flow monitor FM cache format table include layer4 tcp-flags ipv4 sour dest prot tos count pack byte location 0/0/CPU0


Mon Apr 19 09:31:19.589 EDT
Cache summary for Flow Monitor FM:
Cache size:                          10000
Current entries:                         1
High Watermark:                       9500
Flows added:                             1
Flows not added:                         0
Ager Polls:                            580
  - Active timeout                       0
  - Inactive timeout                     0
  - TCP FIN flag                         0
  - Watermark aged                       0
  - Emergency aged                       0
  - Counter wrap aged                    0
  - Total                                0
Periodic export:
  - Counter wrap                         0
  - TCP FIN flag                         0
Flows exported                           0


IPV4SrcAddr      IPV4DstAddr      IPV4Prot IPV4TOS  L4TCPFlags   ByteCount    PacketCount         tcp      0        S|           4282560      71     376


Matching entries:                        1


The exporter (MAP)


Export occurs when data in the cache is removed which can occur in one of three ways.


  • Inactive timer expiry : The cache entry expires due to not matching an incoming packet for a specified amount of time. Default value is 15 seconds.
  • Active timer expiry : The cache entry, though still matching incoming packets, has been in the cache so long that it exceeds active timer. Default value is 30 minutes
  • Cache Exhaustion : The cache becomes full, so some of the oldest entries are purged to make room for new entries.


The netflow exporter can be in a VRF, but can not be out of the Mgmt Interface.

Here’s why. The netflow runs off of the line card (LC interfaces and NP) and there is, by default, no forwarding between the LCs and the management Ethernet.

This because the MGMT ether is designated out of band by LPTS (local packet 
transport services). More detail in the ASR9000 Local packet transport services 
document here on support forums).


Netflow records can be exported to any destination that may or may not be local to the LC where netflow is running. For example, LC in slot 1 & 2 are running netflow & the exporter may be connected to an interface reachble via LC in slot 3.

A total of 8 exporters per MAP is allowed.


Troubleshooting Netflow


RP/0/RSP0/CPU0:A9K-TOP#show flow exporter FE location 0/0/CPU0
Tue Nov  16 11:23:41.437 EST
Flow Exporter: FE
Flow Exporter memory usage:  3280812
Used by flow monitors: FM




Status: Disabled  ><<<<<<<><><><><><>>>>>>>>>><<<<<<<<<
Transport   UDP
Destination         (1963)
Source         (12162) <<<<<<< PROBLEM!!
Flows  exported:                                   0 (0 bytes)
Flows  dropped:                                    0 (0 bytes)




Exporter fields

Export protocol only supported is Netflow v9 over UDP.

Flow record key fields

  • IPv4 source & destination addresses
  • IPv4 protocol & TOS byte
  • Transport source & destination ports
  • Input interface
  • Direction
  • Forwarding status
  • BGP next-hop


Flow record non-key fields

  • TCP flags
  • IPv4 source & destination prefix mask length
  • Source & destination BGP AS #
  • IPv4 address of BGP next hop
  • Packet & byte counters
  • First & last packet timestamps
  • Output interface


Not supported features

Ø Export format v5, v8


Ø FEC for all types of labels
Ø Netflow activation on per MQC class basis
Ø NetFlow for L2 switched traffic

Ø Cisco netflow mib is not supported.



Netflow cache size considerations

Cache size too small : If the number of flows is significantly larger than the cache, the router will perpetually age records and consume more CPU time and bandwidth to export.
Cache size too large : It can waste memory if the number of flows is too less compared to cache size.
Trident ASR9K LC, even with max cache size of 1M, the memory consumption increases by 2.5% per ipv4 FM. (116 MB from 4GB)
Typhoon ASR9K LC, even with max cache size of 1M, the memory consumption increases by 1.25% per ipv4 FM. (116 MB from 8GB)

Recent enhancements

A few nice enhancements have been added to netflow recently:
- ability to tune the rate-limiting for cache expiration (see above for more detail)
- bgp attribute download for full AS-PATH and communities (need to configure bgp attribute download)
- visibility on the phyiscal member used for this record on a bundle or mpls TE tunnel (needs outphyint config on the monitor map)
- ability to see which interfaces are using this flow monitor map. this is handy in case you need to modify the netflow config and have a list of all intfs using the map.
- and yes :) we are working on an inline modification too :)

Show commands

show flow exporter-map ..

show flow monitor-map ..

show sampler-map ..

show processes memory location <0/0/CPU0> | inc nfsvr

show flow monitor .. cache internal location <0/0/CPU0>

show flow exporter .. location <0/1/CPU0>

show flow platform producer statistics location <0/0/CPU0>



show flow platform nfea policer np <np_num> loc <node-id>

show controller np ports all location <0/Y/CPU0>

show controller np count np<number> loc <0/Y/CPU0>

Community Member

Hi Xander,

I am wondering if there were some changes regarding Netflow config. requirements.  I  am observing and analyzing the netflow stats of the ASR9006, Cisco IOS XR Software, Version 4.3.4.  Netflow source is the IP address of Loopback1, sampler  “1 of 10000” and the traffic to the monitoring system is routed via mgmt interface. Currently only 1/10000 of traffic (stacked protocol) is seen at the monitoring system.  Is it caused by Netflow traffic being routed via mgmt interface?


Cisco Employee

hi there,

routing out the mgmt interface is not supported, this because XR will block fabric forwarding through the mgmt interface, that you can override, but won't recommend it.

IF you set the 1/10000 then obviously you wont see every (short lived) flow, but if the flow is in the cache it should properly get exported also. You may want to check the flow exporter stats to see if records are getting dropped.



Community Member

Hi Xander,


At this point I don’t find records being dropped but the monitoring system doesn’t support sampling rates.  What performance increases to expect by setting the sampler to “ 1 out of 1”?





Cisco Employee

I would highly recommend a flow collector that can handle sampling as 1:1 at 10G rates will result in a lot of flow records that will overwhelm your collector at some point.

Running 1:1 on a9k is not an issue as long as your rates are within the set limits. that is 200k pps for typhoon, when there is only one npu that is running netflow.

If the rate exceeds that value, the NPU will replicate the packet for that 1:1 sampling, but it can't punt it to the LC CPU because LPTS will limit it. That will cause also inaccuracies in your flow data.

Same deal with the LC export, if the flow records are sent and not received becaue your collector cant handle it, your info will be inaccurate also.

So it can be done, but there are some dependencies.



Mikkel Troest

Hi Xander

Two questions:

We see enormous figures in the "Flows not added" counter in "show flow monitor cache..." In one scenario it is on a 1:1 sampling on a 10G interface (our own fault ;-) but the count is much, much bigger than the count in the punt-policer (also "not added" increases even at low packet rates, where the punt policer does not kick in)

The other scenario (and different customer) where we see high "not added" numbers is in a 1:00 sampling on a 10G interface, with a pps at around 2500.

What can be causing this? - "Emergency Exports" is zero in both scenarios.


The other Q:

Is there any reason why we don't have custom flow records like in FNF on IOS - or any other form of flow aggregation on the A9K?

If it's in the works, how far down the road?



Cisco Employee

hi mikkel, the flows not added is seen when the cache nears its maximum size and the flows could not be added as the entries were not freed in time by aggressive export.

recommend tuning the cache size to mitigate this.

FNF is coming in I want to say 54 (need to reconfirm with marketing folks), it is definitely on the sw roadmap.



Community Member

Hi Xander,

We run into an issue using the address defined at the loopback interface as exporter source while using the mgmt interface for day to day operational management.

The problem is due to the fact our flow collector does use the source address of the flow packets also to collect additional data via snmp from the router (like interface descriptions). The latter is only allowed via the mgmt interface....  Using different addresses for the same device for snmp and data collection is not supported by the collector.

So the question is, is there a way around this (besides moving all mgmt stuff out of mgmt vrf towards the main routing plane...).

regards, Andre

Cisco Employee

hi andre,

you can try to make this happen with

RP/0/RSP0/CPU0:A9K-BNG(config)#rp mgmtethernet forwarding

and defining a loopback on your system and set the exporter source to that and have the routing take care of reaching the netflow server via the mgmt ethernet.

Although I recommend against it, because this command I suggest is effectively taking the batteries out of the smoke detector and allows full routing between fabric/linecard-interfaces and mgmt ethernet and effectively defeats the purpose of what LPTS out of band tried to provide you :)




Community Member

Hi Xander,

Under IOS there is an "export-stats" template option (v9) configured with:  "ip flow-export template options export-stats". I was expecting the IOS-XR command for this somewhere under "flow exporter-map <map> => version v9". But unfortunately I was not able to find it.

So is there I way to export the statistics at IOS-XR and if how ?

regards, Andre

Hi Xander
Cisco Employee

hi andre,

the statistics should be part of the template, note that using a permanent cache entries will automatically change the options template also.

this you can configure on the times it should export via flow monitor-map <name> version v9 template <...>




Xander, is netflow supported on BVI interface using Trident cards?

Cisco Employee

unfortunately not.

you could enable it on your core interface(s) to verify the traffic through and from BVI to and from core, but not between EFPs like that.



Is there any way you can think of to enable it on the core and still filter it to BVI ingress/egress traffic? I only want relevant traffic.

Cisco Employee

ah, you're hinting on FNF (flexible netflow), the ability to use an ACL to define what is interesting traffic to be sampled.

this is not there today, XR6.

so right now, you'd have to sample on everything and then on your collector filter out the stuff you're not interested in...

so FNF is your ultimate solution for that, which is coming.


Alexandr Gurbo

Hello Alexander,
How to enable netflow for many interfaces at once?
I have PE router with a lot subinterfaces on Bundle-Ether. This Bundle-Ether is part of MC-LAG configuration with double tag. Count of the subinterfaces all time in increase. Periodically I need to see ingress traffic to Bundle-Ether from all subinterfaces.


Content for Community-Ad