nanog mailing list archives

Re: Aggregate traffic management


From: Serge Maskalik <serge () netvmg com>
Date: Wed, 29 Jan 2003 07:46:03 +0000


   Stanislav, 

  It depends what control mechanism you are using: 

   o routes learned via an IGP - ECMP would work and if it's a single 
     destination host, per-packet loadbalancing between the outgoing 
     links is your only practical choice; rest of ECMP schemes work 
     by distributing flows or routes amongst links

   o learned via BGP and the traffic consists of a variety of flows 
     that all use the same reachability information (BGP route); you 
     could de-aggregate the announcement locally if you have an idea 
     how the per-flow volume maps into the route; BGP Multipath feature
     set exists in most router implementations, but the distribution 
     methods are statistically different

  For the latter, several systems exist in the market place that try 
  to automate TE for BGP-learned routes, one of which is ours. These 
  system require a closed feedback loop for traffic volume per flow 
  and link mappings; this needs to occur close to real time to be 
  effective. 
  
        - Serge  

Thus spake Stanislav Rost (stanrost () lcs mit edu):


Dear NANOGers,

I have a very hands-on question:
Suppose I am a network operator for a decent-sized ISP, and I decide
that I want to "divide" aggregate traffic flowing through a router
toward some destination, in order to then send some of it through one
route and the remainder through another route.  Thus, I desire to
enforce some traffic engineering decision.

How would I be able to accomplish this "division"?  What technologies
(even if vendor-specific) would I use?  

I can think of some methods like prefix-matching classification and
ECMP, but I am still not sure exactly how the latter works in practice
(at the router level) and how one may set them up to achieve such
load-sharing.

Thank you for your expertise and lore,

-- 
Stanislav Rost <stanrost () lcs mit edu>
Laboratory for Computer Science, MIT


Current thread: