[nsp] eBGP routes not balancing

[Danny McPherson]

Reading Jeff's initial post and this entire thread, I suspect this
is precisely what was causing him grief -- I.e., the fact that IOS's
default temporal BGP path selection behavior was functioning just as
the original DE intended -- by attempting to maintain some semblance
of control and data plane stability in the face of route flaps
and the like (note: it was coded before BGP route flap damping).

Fortunately, things like BGP Route Refresh, incrementally updatable
prefix-lists, cleaner implementations and BGP route flap damping have
helped calm some of the churn this temporal route selection behavior
was meant to address.

On Jan 5, 2004, at 4:28 PM, Roger wrote:
>
> page 55 of the pdf talks about how the best route is always
> reselected based upon the age of the peer - and how to change
> that

Without actually downloading the document, I suspect Phil says something
akin to "Age of the BGP RIB Entry", not "peer" or "session  
establishment" age.

> and
>
> http://www.cisco.com/en/US/tech/tk365/tk80/ 
> technologies_tech_note09186a0080094925.shtml
>
> This cisco url talks about how *bgp deterministic-med* and the *bgp  
> always-compare-med
> differ..  This page also mentions that
>
> "**Note:* Cisco Systems recommends enabling the *bgp deterministic  
> med* command in
> all new network rollouts. For existing networks, the command must  
> either be deployed
> on all routers at the same time, or incrementally, with care to avoid  
> possible
> internal BGP (iBGP) routing loops."

You should be especially wary of this if you've got a multi-vendor  
network,
as many vendors (e.g., J) perform deterministic (as opposed to  
temporal) BGP
path selection as a default behavior.  [Also, note that the result  
would be
forwarding loops, NOT routing _information loops :-)]

> lastly the bgp bestpath algorithum - how routes get moved from the bgp  
> rib to the
> routing table.  Also discusses on what commands can be used and how  
> they effect this
> behavior.  ie - skip or change default behaior is paramter xyz is  
> set..  Very helpful
> and full of links...  Also briefly mentions BGP multipath.  Which is  
> kinda neat for
> load balancing purposes but hard to achive because so many attributes  
> need be the same.
>
> http://www.cisco.com/en/US/tech/tk365/tk80/ 
> technologies_tech_note09186a0080094431.shtml

I'd definitely recommend folks enable deterministic-med.  In addition,  
I'd suggest
filtering/reseting MEDs altogether -- unless you're especially keen on  
using them
and fully aware of the associated implications and caveats.

While we're on the topic of MEDs, many of the folks here may be  
interested
in this:

http://www.ietf.org/internet-drafts/draft-ietf-grow-bgp-med- 
considerations-00.txt

Feedback welcome!

-danny