Re: [Roll] Loop free local DODAG repair solution

[Philip Levis <[e-mail]>]

So it's basically a multihop DIS? How is that then a local repair? In the case of a subtree becoming disconnected due to the top of the subtree having a null parent set, those nodes at the top would issue a DIS with Solicited Information to repair their local routes (next hop). Nodes below them in the DODAG shouldn't have to trigger a local repair.

Put another way, can you describe a concrete case where DIS with Solicited Information can't solve the problem, yet DRQ/DRP can? Or is this purely an efficiency optimization? If the latter, I'd love to see concrete results indicating the savings (from a real LLN, not simulation).

Phil

On Oct 30, 2012, at 9:01 AM, Jianlin Guo wrote:

> Hi Phil,
> 
> Thank you for your valuable comments. There may be one more major case where a RPL node has a null parent set, that is, parent node of a RPL node becomes unreachable. This is a typical case in lossy networks. For example, a mobile object could block wireless link between two RPL nodes.
> 
> The key differences between DIS and DRQ/DRP are:
> 1) A DRQ message can travel multiple hops. # of hops is controlled by a parameter named MH (Maximum Hop) in DRQ message.
> 2) Upon receiving a DRQ message, a neighbor node responds with DRP message only if neighbor node has non-empty parent set and a rank lower than rank of the DRQ transmitting node. Otherwise, it may forward DRQ message to its parent node.
> 3) Receiving of a DRP message guarantees the discovery of a new parent node.
> 4) Value of MH parameter controls local repair. MH can be set to a small value, e.g. 1,2 or 3, so that DRQ message do not travel too far.
> 
> Jianlin
> On 10/29/2012 6:09 PM, Philip Levis wrote:
>> I'd like to take a step back here and discuss MaxRankIncrease and the notion of null parent sets.
>> 
>> The two major cases where a RPL node has a null parent set are:
>>   1) The node has no members in its neighbor set
>>   2) The node has members of its neighbor set, but all of them advertise a Rank of INFINITE_RANK. The major case we're concerned about here is when rule 3 of 8.2.2.4 requires nodes to advertise INFINITE_RANK due to the value of MaxRankIncrease.
>> 
>> I don't think the local repair scheme will fix case 1. So we're really talking about case 2.
>> 
>> MaxRankIncrease emerged in specifying RPL as a way to support both ZigBee-style centralized periodic tree reconstruction triggers (MaxRankIncrease == 0) as well as CTP-style completely distributed operation while simultaneously providing a cap to the count-to-infinity problem. So really, the issue here is when you have a network with a small MaxRankIncrease and don't want to reconstruct the entire tree, there are valid parents that could be used, but a node does not have them in its neighbor/parent set.
>> 
>> Thinking about it this way (am I wrong?), I am a bit confused by the claim that the draft is a local repair. Assuming a node has to follow rule 3 in 8.2.2.4, this is not a repair mechanism, but rather a neighbor discovery mechanism. Could you explain how DRQ/DRP solve a problem that a DIS with a proper Solicited Information option can't solve? This was the original intent of DIS with Solicited Information. Or perhaps I'm not understanding DRQ/DRP correctly?
>> 
>> Phil
>> 
>> On Oct 26, 2012, at 12:18 PM, Jianlin Guo wrote:
>> 
>>> Thank you for the paper. I agree with [1] on that "dismantling of the sub-DAG, rooted at the node doing the rank increase, causes more turmoil in the network than the routing loops themselves".
>>> 
>>> Now, consider a case in which a node's parent set becomes empty. In this case, RPL provides following set of actions:
>>> 
>>> 1) Start its own floating DODAG
>>> 2) Poison the broken path
>>> 3) Trigger a local repair
>>> 
>>> Both 1) and 2) actions will increase packet delivery delay time (which may be not acceptable for some applications) and possibly cause packet dropping due to limited buffer size of a LLN node (which may also be not acceptable for some applications). So, trigger a local repair is a practical option. Our local repair mechanism is designed for this purpose and it does not create any loops.
>>> 
>>> Jianlin
>>> 
>>> On 10/26/2012 12:10 PM, C Chauvenet wrote:
>>>> Le 26 oct. 2012 à 17:36, Jianlin Guo a écrit :
>>>> 
>>>>> We compared performance metrics such as packet delivery rate.
>>>> Ok.
>>>> 
>>>> In general do you have a document about your experiments that you would like to share ?
>>>> I think it could be a good way to defend your mechanism.
>>>> 
>>>> There are 2 sub questions related to your draft :
>>>> 
>>>>  - Is there a strong need for an additional mechanism to prevent loops ? (the HbH header option mentioned by phil is already there).
>>>>  - Is your mechanism the good way to do so (overhead induced, efficiency...)
>>>> 
>>>> As mentioned by Phil, this subject has been previously discussed inside ROLL few years ago, and did not recommend to add such mechanisms.
>>>> 
>>>> For instance, [1] concludes that
>>>> 
>>>> "the turmoil caused
>>>> by dismantling of the sub-DAGs in order to increase ranks
>>>> may be much more than what the routing loops themselves
>>>> will cause. Consequently, the use of such loop avoidance
>>>> mechanism in the operation of a DAG based routing protocol
>>>> can not be universally recommended."
>>>> 
>>>> [1] : http://www.emmanuelbaccelli.org/publications/AINA_2010.pdf
>>>> 
>>>> Best,
>>>> 
>>>> Cédric.
>>>> 
>>>>> On 10/26/2012 11:21 AM, C Chauvenet wrote:
>>>>>> Hi,
>>>>>> Thank you for your answer.
>>>>>> See inline.
>>>>>> 
>>>>>> Le 26 oct. 2012 à 15:27, Jianlin Guo a écrit :
>>>>>> 
>>>>>>> On 10/25/2012 12:06 PM, C Chauvenet wrote:
>>>>>>>> Hi,
>>>>>>>> 
>>>>>>>> Le 25 oct. 2012 à 16:01, Jianlin Guo a écrit :
>>>>>>>> 
>>>>>>>>> Hi Michael,
>>>>>>>>> 
>>>>>>>>> For your first question, draft-clausen-lln-rpl-experiences-04 pointed out that "it can be observed that with current implementations of RPL, such as the ContikiRPL                                   implementation, loops do occur - and, frequently. During the same experiments described in Section 13, a snapshot of the DODAG was made every ten seconds. In 74.14% of the 4114 snapshots, at least one loop was observed".
>>>>>>>> Is it something that you observed in your own deployments ?
>>>>>>>> More specifically : did you find similar results ?
>>>>>>> We observed the occurrence of loops, unfortunately we did not measure the percentage.
>>>>>> So how did you evaluate the benefit of the mechanism that you proposed ?
>>>>>> 
>>>>>> Cédric.
>>>>>> 
>>>>>>>> Best,
>>>>>>>> 
>>>>>>>> Cédric.
>>>>>>>> 
>>>>>>>>> For your second question, further investigation and experiments are needed.
>>>>>>>>> 
>>>>>>>>> Jianlin
>>>>>>>>> 
>>>>>>>>> On 10/25/2012 8:08 AM, Michael Richardson wrote:
>>>>>>>>>> Jianlin Guo <[e-mail]>
>>>>>>>>>>  wrote:
>>>>>>>>>>     JG> Dear ROLL WG members,
>>>>>>>>>> 
>>>>>>>>>>     JG> As we all know that loop is an open issue in RPL. Experiment shows that loop
>>>>>>>>>>     JG> occurs quite often. We have proposed a loop free local DODAG
>>>>>>>>>> 
>>>>>>>>>> Can you quantify "quite often"?
>>>>>>>>>> Do you have any metrics for how often loops occur, and what the cost is
>>>>>>>>>> of their repair?
>>>>>>>>>> 
>>>>>>>>>> I think that the WG would be very very very interested in additional -experiences
>>>>>>>>>> draft, or pointers to papers explaining same, that gives a repeateable
>>>>>>>>>> experiment in which loops are observed.
>>>>>>>>>> 
>>>>>>>>>> 
>>>>>>>>> _______________________________________________
>>>>>>>>> Roll mailing list
>>>>>>>>> [e-mail]
>>>>>>>>> https://www.ietf.org/mailman/listinfo/roll
>>> _______________________________________________
>>> Roll mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/roll
> 
>