[c-nsp] BGP timers

[Justin M. Streiner]

On Tue, 8 Apr 2008, Uddin, Tahir wrote:

> BTW, Mark, what is the lowest you would go within the CORE and the
> lowest on the customer WAN link and are there any resource issues
> (memory, cpu) that are of concern.

That depends on many variables.  One of the things the hold timers were 
designed to do is provide a small time buffer between the loss of IP 
connectivity that would take a BGP session down and the actual tear-down 
of that session.  Setting it too low can thrash the CPU on your router if 
you have an unstable link or some other issue introduces instability into 
your backbone.

I have one private network interconnected to my backbone that has BGP hold 
timers set to 15 seconds with a 5 second keepalive.  All of the links are 
local (in buildings on my campus), all of the routers are managed by me 
and are exchanging a very small number of prefixes (<5).  OSPF was not 
practical or possible for several reasons.

I would not do this with customers, especially customers who are announcing/
receiving more than a handful of prefixes, because the potential for
customer-induced instability to hit the CPU on my PE routers and possible
elsewhere is too high.  If you choose to do this, make sure you have
safeguards in place, such as appropriately tuned flap damping, and an 
appropriate limit on the number of prefixes you will announce or accept 
over the BGP session.

jms