Re: [trill] Parent node shift draft - relationship to resilient trees draft.

[R Parameswaran <[e-mail]>]

Hi Donald,

Please see inline:

On Tue, Dec 27, 2016 at 6:07 PM, Donald Eastlake <[e-mail]> wrote:
> Hi Ramkumar,
>
> On Mon, Nov 28, 2016 at 2:50 AM, R Parameswaran
> <[e-mail]> wrote:
>>
>> With regard to draft-rp-trill-parent-selection-01, this email tries to
>> answer the question posed at the mike at IETF 97 Trill meeting - how does
>> this draft relate to draft-ietf-trill-resilient-trees?
>>
>> This is examined below from two different considerations - overlap and
>> extensibility.
>>
>> In summary, there is not much direct functional overlap between the two drafts
>> that I can see, and draft-rp-trill-parent-selection-01.txt could possibly be
>> extended to handle backup DT calculation for node failure scenarios. Details
>> of these are broken out below:
>>
>> A. Considerations for Overlap:
>>
>> At a high level, draft-ietf-trill-resilient-trees relates to link
>> protection, and
>> defines the notion of a primary distribution tree (DT) and a backup
>> distribution tree,
>> where these trees are intentionally kept link disjoint to the extent possible,
>> and the backup tree is pre-programmed in the hardware, and activated upon
>> failure of the primary distribution tree.
>
> I believe that is correct.
>
>> draft-ietf-trill-resilient-trees considers the following algorithmic approaches
>> to the building the backup distribution tree:
>>
>> 1. Pure operator config for links on the backup DT/manual generation of
>>    affinity sub-TLV - this is very tedious and unlikely to scale or be
>> implemented
>>    in practice, and hence is disregarded in the analysis here.
>>
>> 2. Section 3.2.1.1a: Use of MRT algorithms (which will produce conjugate trees -
>>    link disjoint trees with roots for primary and backup trees that
>> are coincident on the same physical rBridge).
>>
>> 3. Section 3.2.1.1b: Once the primary DT is constructed, the links used in the
>>    primary DT are additively cost re-weighted, and a second SPF is run to
>>    derive the links comprising the backup DT. Affinity sub-TLV is used to
>>    mark links on the back-up DT which are not also on the primary DT. This
>>    approach can handle conjugate trees as well as non-conjugate trees (link
>>    disjoint trees that are rooted at different physical nodes).
>>
>> 4. Section 3.2.2: A variation on the section 3.2.1.1b approach, but without
>>    affinity sub-TLV advertisement. Once the primary DT is constructed, costs
>>    for links on the primary DT are multiplied by a fixed multiplier to prevent
>>    them from being selected in a subsequent SPF run, unless there
>>    is no other choice, and the subsequent SPF yields links on the backup DT.
>>
>> All of the approaches above yield maximally link disjoint trees, when applied
>> as prescribed.
>>
>> Approach 4 above does not seem to use affinity sub-TLVs and instead seems to
>> depend upon a network wide agreement on the alternative tree computation
>> algorithm being used.
>>
>> Approaches 2 and 3 uses affinity sub-TLV on the backup DT, for links that are
>> not already on the primary DT. The primary DT does not appear to use affinity
>> sub-TLVs. Additionally, from an end-to-end perspective the backup DT comes
>> into picture when the primary DT fails (this is effectively true even in the
>> 1+1 protection mechanism and in the local protection case), and then again,
>> only until the primary DT is recalculated. Once the primary DT is
>> recalculated, the backup DT is recalculated as well, and can change
>> corresponding to the new
>> primary DT.
>>
>> draft-ietf-trill-resilient-trees cannot directly prevent/mitigate a
>> parent node shift
>> on the primary DT at a given parent node, and while usage of the affinity
>> sub-TLV on the backup DT might confer a parent affinity on some nodes
>> on the backup DT, these are not necessarily the nodes on which the
>> network operator
>> may want/prefer an explicit parent affinity. Further, the backup DT is
>> only used on
>> a transient basis, from a forwarding perspective.
>>
>> Given the above, I don't see much of a functional overlap between
>> draft-ietf-trill-resilient-trees, and the draft-rp-trill-parent-selection-01
>> draft.
>
> OK.
>
>> The two drafts can probably co-exist (personal opinion) as they have different
>> goals, and solve different problems. Maybe its good to have some text in one
>> of them, explaining the relationship to the other.
>
> Since trill-resilient-trees is much further along in the process, it
> might be better to add any such text to trill-parent-selection.
>

[RP]: Thanks for the prior review and questions - I have picked this
up, I have added
text (largely derived from this email thread) under section 7 with an upload,
posted as draft-rp-trill-parent-selection-02.

regards,

Ramkumar


>> B. Extensibility considerations:
>>
>> As called out in section 3, draft-ietf-trill-resilient-trees focuses on link
>> protection.  However, the draft alludes to the possibility of using backup DTs
>> for node protection, but considers it out of scope, while calling out some
>> problems with doing so.
>>
>> On the other hand, draft-rp-trill-parent-selection-01 does not explicitly
>> define the notion of a backup DT, and focuses on protecting parent child
>> relationships  on the primary DT, but possibly can be extended to define the
>> notion of a backup DT for node protection.
>>
>> In draft-rp-trill-parent-selection-01, if a sibling of the configured parent
>> node (the node on which child affinity was explicitly configured) goes down,
>> there is not much point in computing the backup DT corresponding to the downed
>> node, because the forwarding state on the configured parent, and on other
>> nodes in the network, will not change because of the already imposed affinity
>> sub-TLV, binding the child nodes listed in the affinity sub-TLV to the parent
>> originating the affinity sub-TLV, regardless of the presence or absence of the
>> sibling node.
>>
>> But, implementations may need to take care to not disturb the hardware
>> programming already in place, while the tree computation reconverges to
>> (nearly) the same outcome as the prior computed tree, when the sibling
>> node goes down.
>>
>> However, in the case where the parent node on which child affinity was
>> configured goes down, it makes sense to configure a back-up DT with simply the
>> Trill default parent selection rules, but with the tree calculation now
>> excluding the configured parent node with the child affinity. The
>> backup DT can be
>> pre-programmed, and when the configured parent node is seen to go down,
>> its affinity sub-TLVs can be discarded, and hardware programming can switch to
>> forwarding with the backup DT state.
>>
>> There are other considerations that need to be thought through (e.g
>> what if any other node (not parent, not sibling) goes down - do we
>> need a backup tree?) , but
>> draft-rp-trill-parent-selection-01 can possibly be extended in this
>> direction if there
>> is interest in pursuing this further. In any case, feedback is welcome.
>
> Those are some interesting ideas. We could probably work on the
> after/if the draft is adopted by the WG.
>
> Thanks,
> Donald
> ===============================
>  Donald E. Eastlake 3rd   [phone] (cell)
>  155 Beaver Street, Milford, MA 01757 USA
>  [e-mail]
>
>> thanks,
>> Ramkumar