[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: hard questions: request routing




Mark,

	I intended to provide a summary of the three solutions
which I think are still under consideration as soon as I completely
understand Abbie's proposal. I guess we then have to make a choice
and document it in the appropriate requirements.

Oliver


Mark Day writes:
 > I need a volunteer, probably from the active participants in this thread, to
 > summarize the issues and the state of the conversation so far.  I have a
 > sense that we have wandered into a rathole but I'm willing to be convinced
 > otherwise.
 > 
 > If there's new and relevant material being defined or discussed, I need to
 > understand how we propose to capture it -- in particular, whether some of
 > this needs to go into the requirements drafts (what, and into which drafts?)
 > or whether this should be captured in some new draft (what is it, and who's
 > writing it?)
 > 
 > Thanks,
 > 
 > --Mark
 > 
 > Mark Stuart Day
 > Senior Scientist
 > Cisco Systems
 > +1 (781) 663-8310
 > markday@cisco.com
 > 
 > > -----Original Message-----
 > > From: owner-cdn@ops.ietf.org [mailto:owner-cdn@ops.ietf.org]On Behalf Of
 > > Oliver Spatscheck
 > > Sent: Tuesday, April 03, 2001 11:04 AM
 > > To: Abbie Barbir
 > > Cc: cdn@ops.ietf.org
 > > Subject: RE: hard questions: request routing
 > >
 > >
 > > Abbie Barbir writes:
 > >  >
 > >  >
 > >  > >
 > >  > >
 > >  > > Abbie,
 > >  > >
 > >  > > 	I think we are fighting over terminology at this point.
 > >  > > Let me try to make
 > >  > > an example:
 > >  > >
 > >  > >
 > >  > > If I understand you right the DSP is building a content tree.
 > >  > > The content tree is managed by the authoritative CDN which has
 > >  > > a global view of the content tree.
 > >  > >
 > >  > > For example:
 > >  > >
 > >  > > - CDN A is authoritative for a particular customer
 > >  > > - All CDN's B C D have business arrangements with each other
 > >  > >
 > >  > > So possible content trees are, for example:
 > >  > >
 > >  > > I.)
 > >  > >
 > >  > >   A
 > >  > >   |
 > >  > >   B
 > >  > >   |
 > >  > >   C
 > >  > >   |
 > >  > >   D
 > >  > >
 > >  > > or
 > >  > >
 > >  > > II.)
 > >  > >
 > >  > >    A
 > >  > >    |\
 > >  > >    C B
 > >  > >      |
 > >  > >      D
 > >  > >
 > >  > > or ....
 > >  > >
 > >  > > What I meant with "you are constraining the topology at some
 > >  > > early stage"
 > >  >
 > >  > the constrain is in the size of the tree, number of content hops,.
 > >  >
 > >  >
 > >  > > is that the authoritative DPS has to decide early on if the
 > >  > > content tree has topology I.) or II.). And it also has to
 > >  > > assure that the content tree is actually a tree and not a graph.
 > >  > > For example, the business relationship allows for a
 > >  > > content graph of
 > >  > >
 > >  > > 	+A-B-C-D+
 > >  > > 	|-------|
 > >  > >
 > >  > > which would be disallowed by the authoritative DPS since a
 > >  > > loop does not make
 > >  > > any sense in terms of distributing content.
 > >  > >
 > >  > > Is that true?
 > >  > >
 > >  >
 > >  > Close, but not accurate. The DPS does not determine the
 > > topology. It only
 > >  > associate
 > >  > an encoding point in the cumulative path. The point here, is that any
 > >  > traversal on the
 > >  > tree can be expressed in a unique way.
 > >  >
 > >  > So the DPS does not restrict the topology.
 > >  >
 > >
 > > OK. Who is choosing if topology I.) or II.) is used and when? Can
 > > a content
 > > tree have loops? If no (that's what I am assuming) which system
 > > is preventing
 > > them and how?
 > >
 > >
 > >  > > What I meant "deals not very well with CDN failures" was now that
 > >  > > if the DPS has chosen topology I.) early on a failure of
 > >  > > CDN B will be preventing the use of CDN C and D, while
 > >  > > topology II does still allow the use of CDN C even if CDN B fails.
 > >  > > (I think this is also the direction Hillary's comment is targeting.)
 > >  > >
 > >  > > You could now argue that the DPS would recover from this
 > >  > > failure by choosing another content tree. This is what
 > >  > > I call an atomic update since all participants
 > >  > > have to have the same view of the content tree to
 > >  > > prevent request routing loops (since routing is a path
 > >  > > in the content tree). It actually uses CDN A's CDP as
 > >  > > arbiter to administer that update.
 > >  > >
 > >  > > Does my terminology makes more sense now or am I still off?
 > >  > >
 > >  >
 > >  > The RRS (not DPS) is the entity that determine the direction tree.
 > >  > yes, CDNs that peer content will have the same view of the
 > > content tree.
 > >
 > >
 > > OK. So you are suggesting an arbiter. Right? (Similar to Brad's
 > > proposal.) Brad also suggested other ways on arriving
 > > at the same view of a tree, therefore, I am just trying
 > > to understand how you determine the common view in the presence
 > > of faults and delays.
 > >
 > >  >
 > >  > > If I make sense this solution has similar drawbacks as the
 > >  > > arbiter solution proposed by Brad. I guess we would have
 > >  > > to agree if it is a requirement to reveal all current
 > >  > > business relationships to the authoritative CDN for
 > >  > > a particular piece of content. There was some opposition
 > >  > > to that idea during the IETF meeting.
 > >  > >
 > >  > > Oliver
 > >  > >
 > >  >
 > >  >
 > >  > Here, I disagree, the main point of the appraoch is that the
 > > content tree is
 > >  > viewed (avialable to the CDNs) by
 > >  > the participating CDNs, but the way that the redirection tree
 > > is computed is
 > >  > totaly private.
 > >  >
 > >  > No business realtionships are revealed or any thing else.
 > >  >
 > >
 > > You just mentioned above that everybody has to have the same view of
 > > the content tree. This by definition reveals the relationships
 > > represented in the content tree. This is not necessary, but sufficient.
 > >
 > > Oliver
 > >
 > 
 >