Re: [RAM] The mapping problem: rendezvous points?

[Tony Li <[e-mail]>]

Dave,


> Some common application models that are adversely affected by PULL:


It seems to me that these are really more about stack implementation  
models than about applications.  In any case...


> 1) Client resolves a name to a list of addresses, and tries each one
> (with some timeout since there's no guarantee of response) until a
> connection succeeds.  Here if the first packet(s) to an address fail
> (or take too long, but failure happens in PULL schemes once a queue
> overflows) then it fails over to the next address.  As a result it  
> will
> connect via a different path to a different server (and see possibly
> different results, either quantitatively or qualitatively).


Differing paths already happen today in the case of hosts with  
multiple A records or load balancing kit.  The truly problematic part  
about this is that with multiple possible mapping results, sequential  
testing can lead to l
a large initial delay.


> 2) Client resolves a name to a list of addresses, and tries each
> one in parallel and chooses the first one to succeed.  Here a
> deterministic delay (say a closer address is un-cached, but one
> far away is cached for some other reason), again causes the  
> inefficient
> one to be chosen.


This could be alleviated by having the traffic deferred until either  
a sufficient quantity of information had arrived, some heuristic  
threshold for 'good' information had been reached, or if later  
updates caused rerouting.


> 3) Server responds to a simple client request via an asymmetric path.
> No DNS occurs here, and the response goes via a router not involved
> in the client-to-server direction.  A cache miss causes a loss
> (since there's a huge number of such clients, a non-negligible
> percentage will be dropped since you can't queue all the packets
> waiting for the delay).  As a result, the client never gets the  
> response
> and either gives up or fails over to a different server (which might
> have the same problem...)


This could be eliminated by moving the mapping client to the host.

Tony


_______________________________________________
RAM mailing list
[e-mail]
https://www1.ietf.org/mailman/listinfo/ram