Hi Inder -

Some rebuttal to your GlobalID FEdEx analogy inline.   I am firmly convinced that the current GID specifications are superfluous and should be removed or at a minimum converted to the "Options" field (see below).

Now read on to see how brainwashed you have been by the corporate FedEx Borg...  :-)

Best regards
Jerry
On 12/6/12 10:22 AM, Inder Monga wrote:

Let me point to Fedex or UPS as an example, a scalable system of packet delivery end-to-end.

a) Scenario with per-hop connection ID as the only way being suggested: 
If the sending customer, say John, sends a package, he will get a tracking ID. Every time a logical point to point delivery happens, a new ID is generated. so when the Fedex store sends the package to the local warehouse, the local warehouse generates a new ID, and then the plane flies to another town and package enters an intermediate warehouse, a new ID is generated. Lets assume to make this case consistent with NSI, that there is not one company, and each leg is managed by a different administrative entity.
If John gives his package to FedEX,...is John then responsible for going to the warehouse to get a status, and then to the airfreight depot to get a status?  and then to DHL to see if FedEx used them for a portion of the trip? and then to UPS? and then to Redball Express, and then try to discover which carries in Europe or Asia those guys then handed it off to?   No.  Users would get pissed off if FedEx just handed the package to another carrier and washed their hands of it.   But this is your example.  Your example is an intra-domain example.  FedEx uses a single number inside their domain only.  Despite their marketing name for it as a "Global Tracking Number" they have a single administrative traking system and all their systems use it.   And I bet they had to com eup with an NSI analog to track shipments acorss different carriers...and those otehr carriers do not use FedEx tracking numbers in their systems. 

 For a user to discover how his package is being shipped - and which carriers were subcontracted to do it - the user *still* needs to go to FedEx to get the information - and FedEx walks the tree to discover who they allocated it to.   And even for their subs - they only get the status their subs wish to release based upon the FedEx id they used.

FedEx acts as the query point for status as well as the point of service - and indeed uses its own Tracking number to query independent subcontractors or its own internal status.  This is exactly what NSI does in the Query() function.  But those subcontractors do not use FedEx numbers for tracking their own internal operations.    You look at a package and you will see multple tracking information.   THis is largely simplified for FedEx and UPS because they are essnetially a centrally managed domain within themselves across the globe.   They are a google for shipping. :-)   But they are not universal and the FedEx "Global tracking number" is more a marketing name - its not a globally unique identifier - except for FedEx.

For the customer to track, he has to
    1) Get the topology of the entire packet delivery, assuming he has authorization. That is a chicken and egg problem because he does not know which domains or service providers his packet is going to go through as Fedex may have many sub-service providers. He can spend months trying to figure that out, but maybe his local store can query and try to figure it out for him
Exactly - so the user queries the service provider he gave the package to - FedEx, and FedEx walks the tree.  Just like NSI.  
    2) Then he has to send a query to try to get the tracking ID from each of those service providers.
No - Customer does not do this - Fedex does this.  Just like NSI - you send a Query() to the head PA and that agent queries down the service tree recursively and rolls up the result.  One stop shopping.
    3) and then he sends a query to each service provider to find out if it left origin point of that service provider to the end point of that service provider.
    4) and then maybe he finds out where the delivery is at, if it did get delivered and where the problem could have been if it didnt.
Sigh.

When is the last time you told FedEx which transport carriers to use to ship your package?   You just tell them where you want it to go.   They decide how to get it there.  And they assume responsibility for getting it there.  Which is why you can then go to *them* to get a valid status.   If you do not ask them to go end to end, then they will not give you end to end status.   If you ask FedEx to just deliver a package to an exporter in New York, and tell the exporter to deliver it to Copenhagen,, FedEx will have nothing to do with the status of NYC to CPH.   Not their issue.  Likewise with NSI.   If you give the package to FedEX, you go back to FedEx to ask for the status - and FedEx goes out and finds the status from however they arranged for it and  rolls it up and presents it to you - even if they delegated delivery to another carrier.  

If you got a subcontractor's tracking number and asked the sub for a status - they could tell you it came from FedEx, even give you a FedEx number,  and they can probably even push the query up the stack and query FedEx to get the status of the fedEx segment and roll it all up and present it to you.   And this is what the detailed Query() does.  (Except the query up doesn't work in NSI because we have this crazy replyTo bogsity as our sole means to send messages up the service tree.)

You could not use the FedEx tracking number on the UPS system, or DHL, or GOD, or any other.    Nor could you use these other carriers tracking numbers with FedEx. 

This is exactly what we should do with NSI - Let each NSA PA assign their own ConnectionID/ReservationID, and let the NSA PA decide how to subcontract it.  The RA and PA each assign and exchange their own ConnectionIDs as part of the Reserve/Confirm process.  And then if the user wants a status - he takes the ConnectionID that he is given and the NSA that assigned it, and queries that NSA for the end to end path.   No GlobalID required. 

You still need to discover the path it took.  You need more than a globally unique ID to do this.  You need to know which NSA(s) are on the path - the most obvious one is the one that the RA used to start the process.   So now we need to encode NSA information in the GID so you know where to start.   So even if the GID contains that first NSA in the GID, you *still* need to go query that NSA to get a status and the path.    Does ESnet want to honor queries from NSAs that it would not otherwise have honored a Reservation Request?   If you only honor requests from "trusted" NSAs, then remote unknonw NSAs have no choice but to walk the tree to let each NSA that successfully estabslihed the reservation ask for the query.

There is no guaranty that the Global ID you have was actualy passed down to the NSA you are querying - even if you know for fact that connection transits that infrastructure.   Virtualization can hide this information, or the GID was not replicated down - maybe because the PA used different credentials to progress the connection.   Lots of legitimate reasons why the GID is not found.  And lots of ways to screw with the system (for instance - what if I deliberately send several requests that use the same GID...it doesn't even have to be one from my own namespace as it could be one I am simply replicating from a parent RA... )  We have no way of validating or verifying the vercity of the global id, so we are just sending jibberish for all we know.   The one identifier we know is valid is the one each NSA tells us is the segment ID for their portion.  So we recurse down the service tree walking these ConnectionIDs and construct a valid picture of the path and state.


b) Scenario with a Global Connection ID , the way tracking really happens

The sending customer delivers his package to a store and gets a global connection ID.
No.  He gets a tracking ID issued by the particular Carrier that he gave the package to.   The Tracking ID is only "global" in the sense that the particular carrier uses that tracking number to track a shipment *they* are delivering anywhere in the world.   It is not used across *all* carrier's systems, nor is it "globally unique" across carriers.

He comes home and can query against that.
He can only usefully query that tracking number against the carrier that took the package.  No one else.  

He can share the global connection ID to whomever he authorizes and they can get status on the package,
Third party requesters can still only query the specific Carrier you gave the package to if they wish to get a status. A tracking number by itself does not indicate the carrier.    You give the tracking number to a different carrier - they won't know it from Peter.   So anyone with a tracking number and the carrier that issued it can indeed get a status - from that carrier.   This is not a function of a globally unique identifier - this is an authorization policy of that carrier.   A different carrier may require you to login to the account that shipped the package.   And if you gave that number to a friend without telling them which carrier it belongs to, your friend is going to go off on a random exhaustive search poking that number at every carrier he knows about to see if it happens to return a result...and hopefully it is not duplicated in another carrier's system and gives him bogus results.  Now think if you had a rogue automated agent that was just probing for ESnet circuits in order to tee up a cyber attack...like the US did to Iran...

On the other hand, if an arbitrary agent knows a Connection ID (not a GID) and the NSA that issued that Conenction ID, and wishes to discover where that connection originates, what rights should that agent have to go and stalk that connection, or the user who owns it?   Should any agent be able to see how much traffic that connection is carrying?  Should any agent be able to freely find out if it was explicitly asked to transit ESnet - and allowed to do so? and what path it tok thru ESnet?  Or if it explicitly avoided CERnet or some other network?   My contention is that these are *local policy decisions* and should not be backdoored simply because we have legacy software that breaks if we don't grandfather in a weak mechanism that was used in the past.   If ESnet wants to let people see their circiuts, then lower the security profiles you apply. i.e. unlock the door and prop it open, don't make a big hole in the wall next to it. 
third party services like Jeroens don't need special authorization for every package sent and from every customer (it works now in AutoGOLE because there is no security).  It is his package. but in order to cancel it he has to provide authorization that only he can (a secure token or cert based authentication)
Sure - we can have different authorization rules for reading vs writing/modifying status.. we just define different policy rules.  Some can be very lax while others very tight.   But we still need to have authorization decision points that enforce policy.

But what do you mean "special" authorization.   If he requested a service instance, he should be able to query its status with his credentials.  If the carrier decides to use internal private policy to make path decisions and potentially uses different credentials to make it happen, does Jeroen still have rights to know those decisions?   I would contend he does not.  Authorizing a user's request to transport data at a certain rate, does not implicitly mean that the user is entitled to know other aspects about that service that were not explicitly part of the request.     At least not by default.  Such access needs to be authorized by a policy - even if that policy says anyone can do this.     The service was a data transport request - as long as the service is meeting that request, why should access to the engineering details be likewise available?     If someone else wants to query the status of his service instance, the credentials they present should be used to allow or disallow it.  Likewise for canceling the service.  It has nothing to do with who's package it is - it is a function of makng sure the actions performed are authorized - always - by asserting the actor's credentials and the action they wish to perform against a policy rule.   If you lower the authorization level to ...nothing...  Then you are simply saying that you implement a policy where anyone's credentials can perform a any function - you don't bypass the authorization decision point.

And more importantly - the level of authorization required to perform some function is a policy imposed by the administrative body responsible for those resources or services.   Just because *you* in ESnet think that this information should be easily visible does not mean that NORDUnet feels the same way.  And NORDUnet may give you excellent connection service performance and reliability...   And just because NORDUnet might allow any academic user a great deal of latitude in provisioning services across NORDUnet does not mean ESnet will honor a similar policy.  

We can all mutually agree to institute very lax policy, but we should not bypass the policy decision points or the policy enforcement points.  

Every step of the way, the global connection ID is scanned and recorded. Regardless of the administrative provider, there is one consistent ID that can track the package and it is easy for him to know where it is. 
Isn't that a much better service interface?
Nope. FedEx simply calls their tracking a number a "global tracking number".   If I call the NORDUnet Connection ID the "NORDUnet Global Tracking ID" ...would you expect it to work in ESnet?   FedEx can impose their global tracking number on themselves - across all of their international divisions.  But  they can't impose it on the subcontractors.  Even the little guys use their own tracking number for actual operations, and simply correlate the FedEx number with their own.   And they only look for the FedEx "GlobalID" when FedEx comes asking....   If those little guys also also serve UPS, they map the UPS tracking number to their internal system for UPS packages... and track UPS numbers when UPS asks.   And when they are simply tracking their own operations - should they use FedEX Global ID or the UPS Internaional ID?  Or could they use their own number..? 



The reason for the confusion around GlobalConnectionID (and the name can be improved - please suggest and let the group decide), is because it is not mandatory. Because of that the "prototype" NSAs don't fill it out. 

We need to approach the standard not as a mechanism we want to be able to work properly across and beyond the academic networks.  ESnet tends to be focused on the US DoE lab requirements - a relatively limited horizon.  But NSI is intended to work across a wide array of network - R&E networks *AND* commercial networks.  We need to keep this in mind - NSI needs to function in other network contexts - often much MUCH larger and untrustworthy.   And R&E networks should not be less able to secure themselves than commercial networks.  So if we elect to implement a loose authorization policy profile - this is fine.   And I agree can be useful in these development activities.    But we should not define a field to be something we cannot verify.  E.g. if an ID appears in the GID field - what happens if it is *NOT* globally unique? 

Perhaps what we need is an "options" field in every message.  This is akin I guess to what Chin asked about in Oxford but I think that was for just reservation request...  So this field would contain arbitrary type-value pairs (or maybe TLVs.)  The Options field would be mandatory, but could be empty.  The contents are intended to be parameters for non-standard (i.e. implementation specific) functionality.    Since these are optional and non-standard data, the NSI standard should make no rules about how the data elements are defined or how they are handled.  NSI should simply state how they are formatted within a message.  And I mean *NO* rules about their usage - they are non-stanard i.e. "NOT of the standard." (!) so should not be required or referenced by any feature in the standard.  The only thing that should be placed in the standard is that there is a message field that carries 0 or more non-standard parameters.  Since the parameters are non-standard, we can not even stipulate when they ought to be replicated by aggregators.   It is left to each particualr NSA implementations to interpret them or to ignore them altogether.

Thoughts?
Jerry


My $0.02 representing the other side,

Inder





On Thu, Dec 6, 2012 at 4:19 AM, Jerry Sobieski <jerry@nordu.net> wrote:
The GlobalID was created at the request of the perfSONAR developers to enable monitoring - how could you determine which segments belonged to a end-to-end connection without each segment being tagged with a Globally unique identifier?    (IMO this was a situation where old ideas and processes die slowly.)

The proper way to do this is to just use the ConnectionID and walk the tree - a detailed Query().   The detailed Query() exposes the actual domains along the path that are actually responsible for processing the service request - rather than trying to find just the domains that are carrying the bits (think about virtualized transit domains).   The Query() process provides authorized access to this information throughout the service tree, and it provides genuine and reliable associations between the ConnectionIDs assigned at each NSA for its children.   Using GLobalID - without descending the tree - is a scattershot and unreliable means of discovering which segments comprise a reservation, and it bypasses all security policy asserted above.  Bad bad bad.  The Query() command is fully authorized and so completing a detailed query provides all the information about the connection segments belonging to a reservation and does so according to authorization policy along the tree and according to the requesters that actually built (and are paying for and/or are responsible for) the reservation.

The only other application of GlobalID that I recall was a batch query possibility (not currently implemented) - the ability to obtain information on a whole batch of unrelated connection segments.  I.e. if my agent wanted to know about all the connections your NSA was servicing then a global ID might provide some correlation with upstream/downstream segments obtained similarly.   Either way, this poses too many security or privacy issues to count...

So our compromise was to allow a tag assigned by a RA to be carried along the segmentation tree so that Connections that wanted to be monitored by perfSONAR could be - but it was optional.    For perfSONAR monitoring purposes - this tag needed to be globally unique - thus its moniker.

As Vagellis observes - if each NSA in the tree wishes to assert their own GlobalID downward, then the GlobalID from above gets overwritten, or ignored altogether.  (I don't recall - it may be the case that if the GID is present, it was supposed to be replicated down - but this then prevents intermediate aggregator RAs from asserting a GlobalID themselves. )   If we want all GLobalIDs to be carried downward, the field must be able to carry multiple GlobalIDs, possibly from each network along a path (think hop-by-hop "chain" provisioning).  And if the field can carry multiple GLobalIDs, why limit each NSA to only inserting a single globalID?  Why not let each NSA insert multiple tags?   Indeed, is there really any requirement for the tag to be globally unique in some way?     It was only perfSONAR that required a globally unique identifier.

Now two years later, I don't know that there has been any perfSONAR tools modified to actually monitor NSI connections, much less to utilize the GlobalID field to do so.    Does any one know?    Are there any other tools that expect/require the GID to be present? or to be globally unique?  (and how easy will it be to break those tools? :-)

Given that NSI has the detailed Query() function that reliably and securely exposes the end-to-end segmentation, I do not think the GlobalID field is required any more.  Does anyone have continuing/additional reasons for having it?  

Unless there is a specific and compelling reason to retain the field as a general purpose communications field preserved in the Reservation record, I propose we should deprecate it and simplify the protocol.   As a general practice for standards, we should *NOT* retain it simply "just in case"...to retain it we need a specific and compelling reason to include it that makes it worthwhile for every implementation to support it.  "optional requirements" is a oxymoron.  Eitehr we need it, or we do not.    If we retain the field, then we need to redefine it so that it can carry multiple tags and so those tags can be recognized/parsed by generalized other agents, and the field should be constrained in size (don't want MPG4 movies being carried in the signaling messages.)

One last indirectly related note:  The ability of Query() to serve generalized agents -besides the uRA- will rely upon these agents issuing a detailed Query() to an NSA that is NOT the first hop PA at the root of the service tree.  For instance, a NORDUnet perfSONAR agent will see a local Connection ID it needs to monitor - how does it discover which NSA is the first hop?  And should it need to?   Why can't the local agent issue a detailed Query() directly to the local NSA PA - and let the local NSA PA walk it *UP and OVER* in the service tree as well as down the service tree?   This is a "flooding" detailed Query().   The Query() functions the same way as it does now, only that the Query() is passed up to the parent RA as well, and the parent/child link from which the Query() came (if it came from a parent/child NSA in this connection's service tree) is pruned from the recursive flooding - it just gets the result.

By tweeking the Query() in this fashion, we enable a much more powerfull ability to manage the reservation.  This will also be an important capability for Notify() functions - an error condition could be flooded to all NSAs in the service tree using a similar flooding model thus being able to notify all NSAs upstream and downstream in the data plane of local interuptions.    Of course, we won't be able to use the current funky WS respondTo process for these upward bound messaging...we'll need a real symmetric session MTL model to do this.   But we already know this.
 
Jerry


On 12/6/12 5:06 AM, Jeroen van der Ham wrote:
Hi,

On 5 Dec 2012, at 18:24, Vangelis Chaniotakis <haniotak@es.net> wrote:


Oh, speaking of globalreservationid!

we're only passing it around and persisting it not doing anything with it AFAICT. 


- it's supposed to be a way to tie this connection with external services, right?
- if so, the name is not quite descriptive of its function
- it's also optional, while the name sounds terribly important
- why a URI instead of a string? or key-value pair  
- there's only one of them, why not allow for a set? 

It is supposed to be the ID for the global reservation. The aggregator NSA receives a request, generates an ID and uses this global ID to make connection requests to each of the participating NSAs. They generate a connection ID for their segment, but must be able to relate that to the global reservation ID.

So yes, it does perform a very important global function :)

It is indeed also meant for tieing in to other services. The URN is used to make it very clear that this is a network global reservation ID.

Jeroen.

_______________________________________________
nsi-wg mailing list
nsi-wg@ogf.org
https://www.ogf.org/mailman/listinfo/nsi-wg


_______________________________________________
nsi-wg mailing list
nsi-wg@ogf.org
https://www.ogf.org/mailman/listinfo/nsi-wg