Re: [Ila] [5gangip] ILA forwaring [Was Re: Problem Statement]

[Tom Herbert <[e-mail]>]

On Tue, May 1, 2018 at 9:58 AM, Joel M. Halpern <[e-mail]> wrote:
> If you do not use caches in your ILA-Ns (and yes, I understand your
> reasoning for not doing so), then you are constructing an overlay network.
> One of the arguments I was given for using ILA was to enable direct
> forwarding of packets effectively without needing to have routing track the
> moving entities.  Without caches, you are pushing all the traffic through
> fewer entities.  And you seem to be either using a lot of ILA-R with
> concomitant information distribution or few ILA-R restricting the
> information distribution problem but instead having traffic concentration
> problems.
>
> I understood from your earlier presentation that the ILA-r using the packet
> as a signal was to avoid dropping the first packet, and as a side-effect not
> needing a separate query message.
> Now you seem to be saying that your think it important to support not having
> caches in the ILA-Ns, which is a VERY different trade-off.
>
Joel,

I think you've misunderstood my position. Caches are _very_ important
to eliminate the cost triangular routing (latency, average path load).
This reduces latency and reduces average load on ILA-Rs. But, and this
is the critical part, caches are only an _optimization_ in ILA. That
means if the cache is rendered ineffective (like by a well crafted DOS
attack) then the only effect is that the optimization is loss (i.e.
greater latency due to triangular router)-- this is quantitively the
worst effect of the attack on an ILA cache. This can be contrasted
that to LISP where the worst case effects of a DOS attack on the cache
is loss of service for users (infinite latency since packets can be
dropped or indefinitely blocked on a cache miss).

Tom

> Yours,
> Joel
>
>
> On 5/1/18 12:37 PM, Tom Herbert wrote:
>>
>> On Tue, May 1, 2018 at 9:10 AM, Joel M. Halpern <[e-mail]>
>> wrote:
>>>
>>> Three reactions, all personal opinion (in case someone thinks my having
>>> helped chair the BoF is in any way relevant; it isn't):
>>>
>>> 1) If ILA-Ns do not have caches, the ILA-Rs will become hot-spots in the
>>> network.  Yes, you have provision for multiple of them sharing load.
>>> However, if that sharing gets to be a significant percentage of teh
>>> routers
>>> in the network, then there is no point in having bothered with ILA, you
>>> are
>>> just routing on the SIR.
>>>
>> Joel,
>>
>> If you provision a network (or any system really) based on an
>> assumption that caches will always attain some hit rate this is a
>> fundamental mistake. One of the goals of a DOS attack would be to
>> drive the hit rate to zero in which case someone will be in a world of
>> hurt. Caches and DOS are a hard mix to contend with in nearly any
>> context, that's why it's much better to view caches as an optimization
>> rather than a requirement. They can be used to alleviate load, but
>> that cannot be relied upon.
>>
>> I would also point out that caches only make sense as internal devices
>> for intra domain communications. This does not make sense for edge
>> routers that would need to create a working set cache for any
>> aribtrary load of traffic from the Internet.
>>
>>> 2) As far as I can tell, when some ILA-N have caches, the ILA-R have no
>>> way
>>> of knowing whether the ILA-N have caches or not.  I can understand what
>>> happens if all ILA-N in a network have the same cache state (either they
>>> all
>>> have caches or they all do not have caches).  But I do not know what
>>> behavior you expect of an ILA-R if the ILA-N are not uniform. Given the
>>> hot-spot issue above, I think you need to really explain why ILA-N would
>>> not
>>> ahve caches.
>>>
>> ILA-Rs and ILA-Ns communicate via ILAMP protocol. That can include
>> capabilities description.
>>
>>> 3 - Minor) Your usage of "sharding" seems odd.  You are simply dividing
>>> the
>>> domain into address blocks, and distributing responsibility for those
>>> blocks
>>> separately.  In other contexts, sharding seems to be used more generally
>>> for
>>> having subcollections of the data which can be moved around.
>>>
>> Sharding is a database term that describes partitioning of the
>> database into smaller chunks for manageablibilty. That is what is
>> happening here (literally in the implementation since we are use a
>> database backend).
>>
>> Tom
>>
>