[rbridge] per-VLAN instances of IS-IS

Radia Perlman Radia.Perlman at sun.com
Wed Jun 20 12:59:56 PDT 2007


I'm also assuming we are doing multipathing, not just of unicast, but 
also multicast.

Radia


Anoop Ghanwani wrote:
> Personally, I would like the group to get to a
> consensus on the ECMP issue fairly quickly.
>
> I too, just like Silvano, thought it was going
> to be done.
>
> Anoop 
>
>   
>> -----Original Message-----
>> From: Eric Gray (LO/EUS) [mailto:eric.gray at ericsson.com] 
>> Sent: Wednesday, June 20, 2007 7:29 AM
>> To: Silvano Gai; Anoop Ghanwani
>> Cc: rbridge at postel.org; Radia Perlman
>> Subject: RE: [rbridge] per-VLAN instances of IS-IS
>>
>> Silvano,
>>
>> 	Yes, that was probably discussed earlier in May.
>>
>> 	AFAIK, discussion != consensus.
>>
>> 	Possibly your customers need to explain to you that the 
>> difference between "core" and "not core" is not always as 
>> clear cut as you apparently think it is.  Nor is it 
>> necessarily going to be simple for a device to determine 
>> which of the available equal cost paths only contain core elements.
>>
>> 	Also, ECMP - at least in the traditional sense - takes 
>> place where equal costs exist in a "routing" domain.  That is 
>> at least as likely to occur in locii containing one or more 
>> edge components, as it is anywhere else.
>>
>> 	There are certainly limited applications - typically 
>> requiring careful configuration to restrict ECMP-like 
>> activity to very specifically defined equal cost paths - 
>> where it will be possible to say "we do ECMP only in the 
>> core."  I personally wonder why link-bundling wouldn't be a 
>> better solution in those cases, however.
>>
>> Thanks!
>>
>> --
>> Eric Gray
>> Principal Engineer
>> Ericsson  
>>
>>     
>>> -----Original Message-----
>>> From: Silvano Gai [mailto:sgai at nuovasystems.com]
>>> Sent: Tuesday, June 19, 2007 11:39 PM
>>> To: Eric Gray (LO/EUS); Anoop Ghanwani
>>> Cc: rbridge at postel.org; Radia Perlman
>>> Subject: RE: [rbridge] per-VLAN instances of IS-IS
>>> Importance: High
>>>
>>> We had this discussion on May 8th, 2007 on this mailing list.
>>>
>>> In a nutshell: we do ECMP only in the core, but we don't 
>>>       
>> learn in the 
>>     
>>> core, so I don't see the problem.
>>>
>>> -- Silvano
>>>
>>>       
>>>> -----Original Message-----
>>>> From: Eric Gray (LO/EUS) [mailto:eric.gray at ericsson.com]
>>>> Sent: Tuesday, June 19, 2007 7:14 AM
>>>> To: Silvano Gai; Anoop Ghanwani
>>>> Cc: rbridge at postel.org; Radia Perlman
>>>> Subject: RE: [rbridge] per-VLAN instances of IS-IS
>>>>
>>>> Silvano,
>>>>
>>>> 	Your response is confusing.  If we are not expected to preserve 
>>>> forwarding symmetry, have we given up on presumed 
>>>>         
>> compatibility with 
>>     
>>>> standard bridges?
>>>>
>>>> --
>>>> Eric Gray
>>>> Principal Engineer
>>>> Ericsson
>>>>
>>>>         
>>>>> -----Original Message-----
>>>>> From: Silvano Gai [mailto:sgai at nuovasystems.com]
>>>>> Sent: Tuesday, June 19, 2007 9:13 AM
>>>>> To: Eric Gray (LO/EUS); Anoop Ghanwani
>>>>> Cc: rbridge at postel.org; Radia Perlman
>>>>> Subject: RE: [rbridge] per-VLAN instances of IS-IS
>>>>> Importance: High
>>>>>
>>>>>
>>>>> Eric,
>>>>>
>>>>> You say: "we're expected to preserve forwarding symmetry".
>>>>>
>>>>> WE ARE NOT.
>>>>>
>>>>> Where does the current draft preserve forwarding symmetry?
>>>>>
>>>>> In regard to your conclusions, many of us (Anoop , Dinesh
>>>>>           
>>> and I for
>>>       
>>>>> sure) are participating because we are building real products.
>>>>>
>>>>> I don't buy your argument that "Very uninteresting
>>>>> (technology) is very
>>>>> good."
>>>>>
>>>>> -- Silvano
>>>>>
>>>>>           
>>>>>> -----Original Message-----
>>>>>> From: rbridge-bounces at postel.org
>>>>>>             
>>> [mailto:rbridge-bounces at postel.org]
>>>       
>>>>> On
>>>>>           
>>>>>> Behalf Of Eric Gray (LO/EUS)
>>>>>> Sent: Tuesday, June 19, 2007 5:27 AM
>>>>>> To: Anoop Ghanwani
>>>>>> Cc: rbridge at postel.org; Radia Perlman
>>>>>> Subject: Re: [rbridge] per-VLAN instances of IS-IS
>>>>>>
>>>>>> Anoop,
>>>>>>
>>>>>> 	I think we strongly disagree on what it means
>>>>>>             
>>> to incorporate
>>>       
>>>>>> something in the control protocol.  From there, it is
>>>>>>             
>>> likely that
>>>       
>>>>>> we also disagree on how best to avoid doing so.
>>>>>>
>>>>>> 	Figuring out how to correctly carry
>>>>>>             
>>> IGMP-snooped information
>>>       
>>>>>> in TRILL protocol is effectively the same thing as 
>>>>>>             
>> incorporating 
>>     
>>>>>> IGMP in the control protocol.
>>>>>>
>>>>>> 	Making a list of everything that bridges snoop
>>>>>>             
>>> and ensuring
>>>       
>>>>>> we can also include that information in IS-IS is
>>>>>>             
>>> exactly the sort
>>>       
>>>>>> of thing we should be trying to avoid.  The good news 
>>>>>>             
>> is that it 
>>     
>>>>>> also should be easy to avoid.
>>>>>>
>>>>>> 	Whatever protocols we might use, TRILL is
>>>>>>             
>>> essentially a layer
>>>       
>>>>>> two forwarding technology.  The messages that layer two
>>>>>>             
>>> forwarding
>>>       
>>>>>> technologies want to snoop are visible to the devices
>>>>>>             
>>> involved in
>>>       
>>>>>> forwarding.  Unless devices actively interfere with
>>>>>>             
>>> "normal" layer
>>>       
>>>>>> two forwarding, "external" control messages should be
>>>>>>             
>>> no different
>>>       
>>>>>> than any other form of data - as far as RBridges are 
>>>>>>             
>> concerned - 
>>     
>>>>>> and should be forwarded in the same way.
>>>>>>
>>>>>> 	We know that we're expected to preserve
>>>>>>             
>>> forwarding symmetry
>>>       
>>>>>> - at least for the layer two encapsulation of TRILL
>>>>>>             
>>> frames. In the
>>>       
>>>>>> same way, and for similar reasons, we may need to 
>>>>>>             
>> preserve path 
>>     
>>>>>> congruency.  This is implicit in the decision to 
>>>>>>             
>> learn from data 
>>     
>>>>>> in the unicast case.
>>>>>>
>>>>>> 	What we don't know is the complete list of
>>>>>>             
>>> things that depend
>>>       
>>>>>> on our doing that.  But we can compose a partial list, and it
>>>>>>             
>>> seems
>>>       
>>>>>> to me that "snooping" (in general) is one example, 
>>>>>>             
>> OA&M may well
>>     
>>> be
>>>       
>>>>>> another and there are bound to be more.
>>>>>>
>>>>>> 	As for the notion of "crippling the technology
>>>>>>             
>>> and making it
>>>       
>>>>>> very uninteresting" - well, to many people, technology becomes
>>>>>>             
>>> very
>>>       
>>>>>> interesting when it is very complicated to make it
>>>>>>             
>>> work.  For the
>>>       
>>>>>> users of any technology, it is only truly useful when 
>>>>>>             
>> it is not
>>     
>>> very
>>>       
>>>>>> interesting.
>>>>>>
>>>>>> 	Paraphrasing something a colleague once told
>>>>>>             
>>> me: a technology
>>>       
>>>>>> is only really useful when the interesting thing in 
>>>>>>             
>> using it is
>>     
>>> not
>>>       
>>>>>> about the technology, but about its use.  Put another way, you
>>>>>>             
>>> know
>>>       
>>>>>> that transportation is useful when the interesting thing about
>>>>>>             
>>> using
>>>       
>>>>>> it is arriving at the destination - rather than the 
>>>>>>             
>> trip itself.
>>     
>>>>>> 	Uninteresting is good.  Very uninteresting is very good.
>>>>>>
>>>>>> --
>>>>>> Eric Gray
>>>>>> Principal Engineer
>>>>>> Ericsson
>>>>>>
>>>>>>             
>>>>>>> -----Original Message-----
>>>>>>> From: Anoop Ghanwani [mailto:anoop at brocade.com]
>>>>>>> Sent: Monday, June 18, 2007 7:08 PM
>>>>>>> To: Eric Gray (LO/EUS)
>>>>>>> Cc: rbridge at postel.org; Radia Perlman
>>>>>>> Subject: RE: [rbridge] per-VLAN instances of IS-IS
>>>>>>> Importance: High
>>>>>>>
>>>>>>>
>>>>>>>               
>>>>>>>> -----Original Message-----
>>>>>>>> From: Eric Gray (LO/EUS) [mailto:eric.gray at ericsson.com]
>>>>>>>> Sent: Monday, June 18, 2007 11:43 AM
>>>>>>>> To: Anoop Ghanwani
>>>>>>>> Cc: rbridge at postel.org; Radia Perlman
>>>>>>>> Subject: RE: [rbridge] per-VLAN instances of IS-IS
>>>>>>>>
>>>>>>>> Anoop,
>>>>>>>>
>>>>>>>> 	I tend to view your response below as yet 
>>>>>>>>                 
>> another reason 
>>     
>>>>>>>> why RBridges MUST use a common path for unicast and 
>>>>>>>> multicast (as well as broadcast and flooded) traffic.  In 
>>>>>>>> fact, it is a very good reason to limit use of
>>>>>>>>                 
>>> "multipathing"
>>>       
>>>>>>>> in general.
>>>>>>>>                 
>>>>>>> If we do place those constraints then it would cripple the 
>>>>>>> technology and make it very uninteresting (at least to me).
>>>>>>>
>>>>>>>               
>>>>>>>> 	It is not clear exactly what you mean by your 
>>>>>>>>                 
>> reference to 
>>     
>>>>>>>> differences between control and data paths.  Control 
>>>>>>>> messages are from one RBridge to another and - 
>>>>>>>>                 
>> presumably - 
>>     
>>>>>>>> follow the same path as data addressed from RBridge to 
>>>>>>>> RBridge.  That path may be different from data
>>>>>>>>                 
>>> transitting an
>>>       
>>>>>>>> RBridge campus, but there is no RBridge "control" traffic 
>>>>>>>> that transits a campus.
>>>>>>>>
>>>>>>>> 	Are you suggesting that IGMP should be treated 
>>>>>>>>                 
>> as a control 
>>     
>>>>>>>> protocol within TRILL?
>>>>>>>>                 
>>>>>>> In this case, when I said control I meant IGMP 
>>>>>>>               
>> (since that is 
>>     
>>>>>>> what is being used to effect pruning).  If we snoop on the 
>>>>>>> information at the edge of the trill cloud and pass this 
>>>>>>> information around in IS-IS, then we don't need to 
>>>>>>>               
>> worry about 
>>     
>>>>>>> treating it as a separate control protocol.
>>>>>>>
>>>>>>>               
>>>>>>>> 	Nor is it very clear why this observation is 
>>>>>>>>                 
>> specific to 
>>     
>>>>>>>> our discussion of multicast optimization.  As a general 
>>>>>>>> rule, if there is one reasonably well know 
>>>>>>>>                 
>> application that 
>>     
>>>>>>>> relies on path congruency for proper operation, there are 
>>>>>>>> probably others.
>>>>>>>>
>>>>>>>> 	Do you think we should provide a logical 
>>>>>>>>                 
>> control-function 
>>     
>>>>>>>> "bridging" service for all such
>>>>>>>>                 
>>> applications
>>>       
>>>>>>>> as a result of path divergences we introduce?  Should we 
>>>>>>>> conduct a research project to determine what 
>>>>>>>>                 
>> other "control 
>>     
>>>>>>>> protocols" we need to include in TRILL?
>>>>>>>>                 
>>>>>>> We would need to make a list of everything that 
>>>>>>>               
>> bridges snoop 
>>     
>>>>>>> on and make sure we have ways of sending that 
>>>>>>>               
>> around in IS-IS 
>>     
>>>>>>> (if we care about those functions).
>>>>>>> I'm not aware of anything other than IGMP that trill would 
>>>>>>> need to worry about, though.
>>>>>>>
>>>>>>> Anoop
>>>>>>>
>>>>>>>               
>>>>>> _______________________________________________
>>>>>> rbridge mailing list
>>>>>> rbridge at postel.org
>>>>>> http://mailman.postel.org/mailman/listinfo/rbridge
>>>>>>             



More information about the rbridge mailing list