Re: [altoext] General Applicability of a Cost Graph?

[Greg Bernstein <[e-mail]>]

Hi Richard, I think we are in agreement. Just a few comments and 
questions below.
Cheers
Greg
On 4/17/2012 8:59 AM, Y. Richard Yang wrote:
> Greg,
>
> Good answers. Please see below.
>
> On 4/17/12 10:33 AM, Greg Bernstein wrote:
>>> I agree that Cost Graph, in addition to Cost Map, is a good concept 
>>> to add to ALTO. Conceptually:
>>>
>>> - (E2E) Cost Map: "An ALTO Cost Map defines Path Costs pairwise 
>>> amongst sets of source and destination Network Locations.  Each Path 
>>> Cost is the end-to-end cost from the source to the destination." The 
>>> key is that the information is end-to-end.
>>>
>>> - (Link) Cost Graph (or Link-Cost Graph):  represents per-link info.
>>>
>>> I see that we need a path-vector (E2E) Cost Map to connect the (E2E) 
>>> Cost Map and the (Link) Cost Map, where the two maps need to be 
>>> defined on the same set of network nodes (Network Map).
>> -->  I think your semantics comment below applies here.
>>
>> I was imagining the leaf nodes in the "Cost Graph" being ALTO PIDs. 
>> Non-leaf nodes and links are abstractions for modeling network 
>> constraints and costs. Hence I figured that at a minimum I would need 
>> an ALTO Network map to map endpoints to PIDs and the Cost Graph.  In 
>> addition, I could use endpoint properties and costs to model the 
>> final "hop" from PID (node) to endpoint.
>>
> This graph representation by introducing internal nodes to represent 
> constraints and costs is quite interesting. I feel that for 
> generality, the internal nodes may not be limited to virtual 
> (abstract) nodes, but can be real PID nodes, representing, for 
> example, loose source routing, locations/PE.
--> Agree. My statement above on"non-leaf" nodes isn't quite what I 
intended. Better "a node without any attached endpoints" is used for 
modeling purposes.
>
>> How would you see a path vector being used by an application?
> A use case of path vector, with each element being a node from the 
> Cost Graph, is to convey the path(s) when there are multiple paths 
> from a source PID to a destination PID on the Cost Graph, and the 
> provider has chosen one or a few candidates. Application could use the 
> path vector for reliability/availability selection (some examples in 
> your presentation): suppose PID_client is a client site, and there are 
> two sites PID_s1, PID_s2 hosting replicated servers. An application 
> may check the path vector of PID_s1 -> PID_client  and that of PID_s2 
> -> PID_client when selecting paths: selecting disjoint paths for 
> reliability, selecting maximum flow paths to maximize the downloading 
> rate of the client, ....
>
--> Makes sense. We would somewhere need a mechanism to allow the 
application to specify which path it intends to use. Right?
So this would be more commonality between the "Data Center" and "High 
bandwidth" uses cases?
> -- snip -- 


-- 
===================================================
Dr Greg Bernstein, Grotto Networking [phone]