The IP Over NBMA (ION) Archive

Cell Relay Retreat>ION Archive>month:1996-Aug> msg00036



[Date Prev][Date Next][Thread Prev][Thread Next]  
  [Date Index][Thread Index][Author Index][Subject Index]

Proposed extensions to Inverse ARP

  • From: SanJeev Rampal <sanjeev@raleigh.ibm.com>
  • Date: Wed, 14 Aug 1996 16:31:50 +22299756

Bryan, Grenville, Paul, thanks for your replies. Some comments follow. 

Folks, I apologise for taking up net bandwidth on this thread and would be 
glad to take it offline whenever requested.  


Cheers, 
Sanjeev Rampal. 

========================================================================
BG> IP addresses to map to a single ATM address. As you point out several
BG> replies would be needed in response to a single request, and then you
BG> have the problem of making this transfer reliable - otherwise some
BG> of the replies could get lost and you would not know.

Reliable transfer is not needed: if some replies are lost, the corresponding
protocol addresses simply don't get attached to the existing connection (in 
the requestor's cache) at this time. Packets to those addresses can always 
take the routed path. 

BG> This seems to imply that the target of a shortcut can learn IP/ATM
BG> mappings from the source which it then uses to forward data, without
BG> actually issuing an NHRP Resolution Request. This is currently frowned
BG> upon by the spec (a SHOULD NOT). One reason is that this may bypass
BG> filtering that would be imposed along the routed path.

Bypassing router filtering on the reverse path is definitely a concern. 
Some points to consider regarding this...

1. This would be an optional feature so that a host/router can 
simply be configured as to whether to respond to these ``extended'' InARP
packets or not. A station which has been configured to reply to 
these InARP packets should be aware of the fact that stations could send 
him packets directly, bypassing any router filtering. 

The point being that there is an upside to this feature (improved 
performance) as well as a downside (bypassing router filtering). 
If it is made an optional feature so that a user/administrator can take 
the conscious decision of selectively turning it on in situations where 
filtering is not required, then it can yield improved performance. 
  
As an example a public web server would want to turn this feature on so 
that the rest of the world can connect with itself very quickly but a 
workstation containing sensitive information would turn it off ...

2. As regards current NHRP wordings, it appears to me that such ``reverse
shortcuts'' can be established even now (maybe someone more knowledgeable 
about NHRP can correct me here). 

a. For instance in section 6.2.1, both for serving NHSs and transit NHSs, 
situations are described where they should/ may cache information from NHRP 
request packets. If such NHSs are also NHCs, clearly they can establish 
``reverse shortcuts'' without going through the NHRP requesting procedure. 

b. Section 2.2 also talks of ``Logical NBMA Subnetworks'' within which 
``unfiltered subnetwork connectivity'' is available. The scope of the extended 
InARP could also be restricted to such administratively configured domains. 

c. Finally, can it actually be verified whether other stations are caching 
source information or not (even if the spec forbids it) ? No. Given that fact, 
some form of administrative partitioning will anyway be needed in all cases 
where security/filtering is an issue.   

BG> If there are things we don't need then we should simplify, rather
BG> than extend, or at least restrict the extensions to PVCs which is
BG> the scope of 1293.

The PVC case implicitly is similar to the configured domain case since 
it is assumed that if an administrator is configuring a connection between 
two endpoints which lie in different LISs, he is fully aware of bypassing 
any router  filtering and is setting up the connection nevertheless. So, yes
it can be allowed for PVCs but why not for SVCs also (as long as its use 
and potential problems are limited using appropriate administrative 
configuration).  

gja> Since this extended form of Inverse ARP is only applicable to
gja> NHRP-enhanced endpoints, why not introduce it as an extension of
gja> the NHRP code space? You'll find that the generic MARS/NHRP fixed
gja> header code point space was chosen to allow this sort of expansion.

Yes it can certainly be realised using NHRP code points instead of ARP 
code points (essentially making it an ``inverse NHRP'' or ``PHRR'' as you 
say).  

pf> I think this is a really bad idea. Primarily, it defeats the purpose
pf> of relying in inverseARP in environments such as frame-relay where
pf> auto mapping of a specific logical subnet to a local identifier
pf> (DLCI) is used. I believe thsi will introduce more problems that it
pf> solves.

Paul, I do not understand your point about the auto mapping of subnets to 
DLCIs. I know very little about Frame Relay so may be missing something. 
I would be interested in learning about this problem if you can give more 
details.