[67932] in North American Network Operators' Group

home help back first fref pref prev next nref lref last post

Re: Converged Networks Threat (Was: Level3 Outage)

daemon@ATHENA.MIT.EDU (Petri Helenius)
Wed Feb 25 14:46:13 2004

Date: Wed, 25 Feb 2004 21:37:24 +0200
From: Petri Helenius <pete@he.iki.fi>
To: David Meyer <dmm@1-4-5.net>
Cc: nanog@merit.edu
In-Reply-To: <20040225191916.GA8564@1-4-5.net>
Errors-To: owner-nanog-outgoing@merit.edu


David Meyer wrote:

>	
>	No doubt. However, the problem is: What constitutes
>	"unnecessary system complexity"? A designed system's
>	robustness comes in part from its complexity. So its not
>	that complexity is inherently bad; rather, it is just
>	that you wind up with extreme sensitivity to outlying
>	events which is exhibited by catastrophic cascading
>	failures if you push a system's complexity past some
>	point; these are the so-called "robust yet fragile"
>	systems (think NE power outage).  
>  
>
I think you hit the nail on the head. I view complexity as diminishing 
returns play. When you increase complexity, the increase does benefit a 
decreasing percentage of the users. A way to manage complexity is 
splitting large systems into smaller pieces and try to make the pieces 
independent enough to survive a failure of neighboring piece. This 
approach exists at least in the marketing materials of many 
telecommunications equipment vendors. The question then becomes, "what 
good is a backbone router without BGP process". So far I havenīt seen a 
router with a disposable entity on interface or peer basis. So if a BGP 
speaker to 10.1.1.1 crashes the system would still be able to maintain 
relationship to 10.2.2.2. Obviously the point of single device 
availability becomes moot if we can figure out a way to route/switch 
around the failed device quickly enough. Today we donīt even have a 
generic IP layer liveness protocol so by default packets will be 
blackholed for a definite duration until a routing protocol starts to 
miss itīs hello packets. (Iīm aware of work towards this goal)

In summary, I feel systems should be designed to run independent in all 
failure modes. If you lose 1-n neighbors the system should be 
self-sufficient on figuring out near-immediately the situation, continue 
working while negotiating with neighbors about the overall picture.

Pete


home help back first fref pref prev next nref lref last post