Re: [ipfix] bytes and packet counters, tstamp of last report

[Benoit Claise <[e-mail]>]

Hi Jeff,

An old email directed to me that I forgot to answer.

> Benoit,
>  
>   One other point on your e-mail.  In your discussion of "Long lived 
> flows", are you saying that if an observed connection lasts say 5 
> minutes, and through configuration the system will export a long lived 
> flow after 3 minutes of activity, that the second flow record's 
> numbytes is measured from the beginning of the flow

> (e..g it represents period t,t+5m vs. t+3m,t+5m, where t is the actual 
> start time of the long lived flow).

The "long lived flow" of your example is divided in 2 flows, hence in 2 
flow records.
The first flow record will have the number of bytes for (t,t+3) and the 
second (t+3,t+5)
When a flow is expired (for whatever reasons), the flow record must 
export the number of bytes of that flow.
To come back to your example, the first flow is expired and exported 
after 3 minutes. A new flow (with exactly the same properties) is 
directly created, but is considered as a independant flow. So cumulative 
data.

>  
>   I would have thought that the second record would indicate flow 
> properties SINCE any previous flow record.  That was the revised 
> wording I had suggested based on Maurizio's observation of ambiguity.
>  
>   If you are saying that the second exported flow for a long lived 
> connection is cumulative, then you've created a rather dicey problem.  
> Especially if a collection system is doing aggregation of these 
> flows.  Because instead of a flow being a discrete self-contained 
> piece of information, one must now determine if it actually represents 
> additional counts added to flows previously reported.  If so this sets 
> you up to have to search an ENORMOUS space of old flows to do delta 
> calculations.
>  
>   My impression of traditional NFvX behavior is that exporting a flow 
> effectively pops it out of the flow table.  Any subsequent traffic for 
> the same flow would first create a new flow entry and the counters 
> effectively reset each time a flow is exported. 

You are fully right!

Regards, Benoit.

> Here is the current Cisco documentation:
>  
>   
> http://www.cisco.com/en/US/products/hw/switches/ps4324/products_configuration_guide_chapter09186a008019d0da.html
>
>     Configuring Netflow Aging Parameters
>      
>     You can control when flows are purged from the software flow cache
>     (and, if configured, reported through NDE) with the configuration
>     aging parameters, Active and Inactive, of the ip flow-cache
>     timeout command.
>
>     Active Aging specifies the period of time in which a flow should
>     be removed from the software flow cache after the flow is created.
>     Generally, this parameter is used to periodically notify external
>     collection devices about active flows. This parameter operates
>     independently of existing traffic on the flow. Active timeout
>     settings tend to be on the order of minutes (default is 30min).
>
>   My reading here is that exporting and purging are one in the same.  
> I.e. I never get a flow which may reflect counts which were already 
> reported in a previous flow.
>  
>   Are you proposing changing this behavior?  If so, this has 
> significant deliterious impact on collection.  I would not recommend 
> changing.
>  
> Regards,
>  
>   Jeff Meyer
>
>     -----Original Message-----
>     *From:* MEYER,JEFFREY D (HP-Cupertino,ex1)
>     *Sent:* Friday, October 10, 2003 11:05 AM
>     *To:* 'Benoit Claise'; MEYER,JEFFREY D (HP-Cupertino,ex1)
>     *Cc:* 'Tal Givoly'; [e-mail]; Maurizio Molina;
>     [e-mail]
>     *Subject:* RE: [ipfix] bytes and packet counters, tstamp of last
>     report
>
>     Benoit,
>      
>       Adding these three information items to the Info Model sounds
>     fine. 
>      
>        What is the rollover and reset behavior of these counters? 
>     E.g. when an observer is started due to reboot or other
>     administrative action the counters are set to zero.  Is there a
>     way to distinguish between a counter reset and rollover?
>      
>       Presumably these should be long's in the information model.  Is
>     there a practical lower bound on an implementation in terms of
>     size, e.g. 32-bits?
>      
>       Also what are the triggers for these to be exported?  Some
>     timer, threshold, or unspecified operation?  In any of these
>     cases, is there a guarantee that there will not be > 1 rollover
>     between exports?
>      
>       I would think most of these aspects should be captured in the
>     information model.  Hopefully they also illustrate some of the
>     problematic behavior of dealing with counters, not insurmountable,
>     but certainly irritating.
>      
>     Regards,
>
>       Jeff Meyer
>
>         -----Original Message-----
>         *From:* Benoit Claise [mailto:[e-mail]]
>         *Sent:* Friday, October 10, 2003 4:11 AM
>         *To:* MEYER,JEFFREY D (HP-Cupertino,ex1)
>         *Cc:* 'Tal Givoly'; [e-mail]; Maurizio Molina;
>         [e-mail]
>         *Subject:* Re: [ipfix] bytes and packet counters, tstamp of
>         last report
>
>         Hi,
>
>>Hi,
>>
>>
>>  In some use cases counters may have desirable properties as Tal
>>has described.  But like many things, counters also have their 
>>limitations and downsides.
>>
>>  In the basic Flow Export scenario which is exemplified by existing
>>Netflow implementations, a flow record indicates a summary of
>>observed behavior.
>>
>>  Consider two flow events:
>>
>>      sourceAddress=1.2.3.4
>>      destinationAddress=1.2.3.5
>>      packetCount=1000
>>      byteCount=100000
>>      flowCreationTime=2003-10-08T10:00:00Z
>>      flowEndTime=2003-10-08T10:00:05Z
>>      TcpControlBits=0x13
>>
>>
>>      sourceAddress=1.2.3.4
>>      destinationAddress=1.2.3.6
>>      packetCount=1000
>>      byteCount=200000
>>      flowCreationTime=2003-10-08T10:00:02Z
>>      flowEndTime=2003-10-08T10:00:06Z
>>      TcpControlBits=0x13
>>
>>  Each of these completely describes a conversation, as it includes both
>>  the fin and syn TCP flags.  Start time, end time and absolute counts.
>>
>>  Given the large percentage of flows which when exported, describe a
>>  complete conversation, I don't see the advantage of using counters.
>>
>>  I'm not even sure how one would propose the use of counters as an
>>  alternative.  Would you require two records be sent for each
>>  flow (one with byteCounter and packetCounter=0 and another with
>>  the final value?)
>>
>>  Since most flows can be represented by a single record, doubling
>>  the number of records to accomodate counters seems a bit odd.
>>  Counters also require the holding of state, which produces
>>  a much greater burden on the collector.  Since the necessary
>>  state must be held by the observer, why double the workload?
>>  I.e. a single   counter is pretty useless.  I need to delta it
>>  against a previously observed matching counter to arrive at a
>>  quantity.
>>
>>
>>  For information models such as PSAMP, there may be a compelling
>>  use case.
>>
>>  So I imagine adding an explicit annotation in the information
>>  model for counters would be worthwile.  However, I think that
>>  counter behavior such as when it is zeroed out and its rollover
>>  behavior need to be appropriately documented and thought out.
>>  (e.g. on restart, the counters presumably reset to zero, so
>>  here is additional state a collector may need to monitor to
>>  reconcile (if possible) existing observations).
>>
>>  For IPFIX's base information model, I would strongly discourage
>>  "fixing" something which isn't broken.
>>
>         I fully agree with Jeff here regarding the flow records but
>         there are actually 2 aspects to this problem.
>
>         1.
>         The flow record.
>         There is an expiration process for the flow records.
>         http://www.ietf.org/internet-drafts/draft-ietf-ipfix-protocol-00.txt
>
>4.1 Flow Expiration 
>    
>   A Flow is considered to be inactive if no packets belonging to the 
>   Flow have been observed at the Observation Point for a given timeout 
>   interval otherwise it is considered as an active flow.  
>   A Flow can be exported under the following conditions: 
>    
>      1. If the Metering Process can detect the end of a Flow, it 
>      SHOULD export the Flow Records at the end of the Flow. For 
>      example, a Flow generated by TCP [3] type of traffic where the 
>      FIN or RST bits indicate the end of the Flow. 
>       
>      2. If the Flow has been inactive for a certain period of time. 
>      This inactivity timeout SHOULD be configurable, with a minimum 
>      value of 0 for an immediate expiration. For example, a Flow 
>      generated by UDP [2] type of traffic. 
>       
>      3. For long-lasting Flows, the Metering Process SHOULD export the 
>      Flow Records on a regular basis. This periodicity SHOULD be 
>      configurable. 
>       
>      4. If the Metering Process experiences internal constraints, a 
>      Flow MAY be forced to expire prematurely (for example, counters 
>      wrapping or low memory). 
>
>What the IPFIX information model tries to say is that: a flow expires and we send the packet/byte count for the flow duration. So this is actually a running counter for the flow duration.
>Note: a long last flow is actually composed a several flows (see condition 3 before). 
>Ok, I understand that this could be better explained in the IPFIX information model draft:
>   The packet count can be a running counter and is the count from the
>   beginning of the flow establishment.
>   The packet count can be a delta counter and is the count since the
>   last report for this flow.
>
>
>http://www.ietf.org/internet-drafts/draft-claise-netflow-9-05.txt speaks of:
>                                            Incoming counter with  
>    IN_BYTES                     1    N     length N x 8 bits for bytes  
>                                            associated with an IP Flow    
> 
>                                            Incoming counter with  
>    IN_PKTS                      2    N     length N x 8 bits for  
>                                            packets associated 
>                                            with an IP Flow  
>
>2. 
>The second aspect is actually total counters.
>Something that needs to added to the IPFIX protocol draft is described in its "open issues" section:
>   - The proposal on the table is to send a IPFIX Sync (this would be 
>   an Options Data Records) message periodically (periodicity is 
>   configurable), with the following information (aside the standard 
>   IPFix header) 
>           * Number of flow records sent (for each template?)  
>           * Packets and bytes sent (for each template?) 
>
>So we need to keep a few running counters.
>
>In http://www.ietf.org/internet-drafts/draft-claise-netflow-9-05.txt, there are explicitely described:
>                                            Counter with length   
>                                            N x 8 bits for bytes 
>    TOTAL_BYTES_EXP              40   N     for the number of bytes  
>                                            exported by the Observation 
>                                            Domain 
> 
>                                            Counter with length  
>                                            N x 8 bits for bytes  
>    TOTAL_EXP_PKTS_SENT          41   N     for the number of packets  
>                                            exported by the Observation  
>                                            Domain  
> 
>                                            Counter with length  
>                                            N x 8 bits for bytes  
>    TOTAL_FLOWS_EXP              42   N     for the number of Flows  
>                                            exported by the Observation  
>                                            Domain  
>      
>
>
>         I propose to also explicitely described them in the IPFIX
>         information model draft.
>
>         Regards, Benoit.
>
>>Regards,
>>
>>  Jeff Meyer
>>
>>_________________________________________________________________
>>Get MSN 8 Dial-up Internet Service FREE for one month.  Limited time offer--
>>
>>sign up now!   http://join.msn.com/?page=dept/dialup
>>
>>  
>>
>>>-----Original Message-----
>>>From: majordomo listserver 
>>>[mailto:[e-mail]]On Behalf
>>>Of Tal Givoly
>>>Sent: Wednesday, October 08, 2003 9:14 AM
>>>To: [e-mail]; Maurizio Molina
>>>Cc: [e-mail]
>>>Subject: RE: [ipfix] bytes and packet counters, tstamp of last report
>>>
>>>
>>>Both counters have value (running and delta). Obviously, they 
>>>must be kept
>>>discrete from an information model perspective as they 
>>>represent different
>>>information. Devices sometimes maintain one, sometimes the other, and
>>>sometimes both. Just as an example, as far as I recall, 
>>>NetFlow v5-8 doesn't
>>>maintain running counters (counter to your suggestion) EVEN THOUGH it
>>>operates over a non-reliable and non-congestion-aware 
>>>transport. Our probe
>>>product emits both for various different purposes.
>>>
>>>Tal
>>>
>>>-----Original Message-----
>>>From: majordomo listserver 
>>>[mailto:[e-mail]]On Behalf
>>>Of Stewart Bryant
>>>Sent: Wednesday, October 08, 2003 8:09 AM
>>>To: Maurizio Molina
>>>Cc: '[e-mail]'
>>>Subject: Re: [ipfix] bytes and packet counters, tstamp of last report
>>>
>>>
>>>
>>>
>>>Maurizio Molina wrote:
>>>
>>>    
>>>
>>>>Hi,
>>>>the IPFIX info model states (sec. 6.10) that
>>>>
>>>>   The packet count can be a running counter and is the 
>>>>      
>>>>
>>>count from the
>>>    
>>>
>>>>  beginning of the flow establishment.
>>>>
>>>>  The packet count can be a delta counter and is the count since the
>>>>  last report for this flow.
>>>>
>>>>      
>>>>
>>>I would prefer that we only supported running counters because:
>>>
>>>a) Supporting two types is more scope for non-interworking, and
>>>    you can always convert from one to the other at the collector.
>>>
>>>b) Running counters also work over an unreliable transport
>>>    whereas delta counters mandate the use of a reliable transport.
>>>
>>>c) Running counters are most likely what the hardware is keeping
>>>    anyway, therefore delta counters are more work and more storage
>>>    at the exporter.
>>>
>>>If we decide that we need both then we have to represent them
>>>as two different information elements because the info model does
>>>not support sub-typing.
>>>
>>>Stewart
>>>
>>>    
>>>
>>>>(There's the same statement for byte counts in 6.11).
>>>>
>>>>To me, this doesn't clarify wheter it is both...and  (2 counters) or
>>>>either .... or (1 counter).
>>>>
>>>>Moreover, there is currently no room for a field containing the
>>>>timestamp of the last report of a flow. I think that keeping this
>>>>timestamp is helpful for many applications and the info model should
>>>>support it.
>>>>Regards,
>>>>Maurizio
>>>>
>>>>
>>>>
>>>>--
>>>>Help        mailto:[e-mail] and say 
>>>>      
>>>>
>>>"help" in message
>>>    
>>>
>>>>body
>>>>Unsubscribe mailto:[e-mail] and say
>>>>"unsubscribe ipfix" in message body
>>>>Archive     http://ipfix.doit.wisc.edu/archive/
>>>>
>>>>
>>>>      
>>>>
>>>--
>>>Help        mailto:[e-mail] and say "help" 
>>>in message
>>>body
>>>Unsubscribe mailto:[e-mail] and say
>>>"unsubscribe ipfix" in message body
>>>Archive     http://ipfix.doit.wisc.edu/archive/
>>>
>>>
>>>--
>>>Help        mailto:[e-mail] and say "help" 
>>>in message body
>>>Unsubscribe mailto:[e-mail] and say
>>>"unsubscribe ipfix" in message body
>>>Archive     http://ipfix.doit.wisc.edu/archive/
>>>
>>>    
>>>
>>
>>--
>>Help        mailto:[e-mail] and say "help" in message body
>>Unsubscribe mailto:[e-mail] and say
>>"unsubscribe ipfix" in message body
>>Archive     http://ipfix.doit.wisc.edu/archive/
>>  
>>
>