[quagga-dev 5968] route cache and slow convergence time on linux

Francesco Dolcini fdolcini at sysnetsistemi.it
Fri Sep 26 16:50:18 BST 2008


I am running ospfd/quagga 0.99.10 on linux (custom embedded 2.4.20x, but 
I think 2.6 would be the same) and doing some preliminary analysis on 
network convergence time.

My current test network is a 4 router loop

              _[R1]_
             /      \
            /        \
           /          \
[H1]----[R4]        [R2]
           \          /
            \        /
             \_[R3]_/


Multipath routing is on, with --enable-multipath=4.

When I break the link between R4 and R1, the time to recover the 
comunication from H1 to R2 is a little more then 3 seconds (measured 
using a ping -i 0.1), 1 second is the minimum I can achieve to detect 
the link failure (ip ospf dead-interval minimal ..., link-detect is 
slower for my specific network interfaces), the other 2 second seems 
related to the default value of /proc/sys/net/ipv4/route/min_delay 
(minimum delay before the route cache is updated or flushed), setting 
this to zero seems to solve this slow convergence issue.

Now the questions:
- Is there any known issue setting this value to zero? the only 
reference I found is this thread on LKML: 
http://www.ussg.indiana.edu/hypermail/linux/net/0010.2/0006.html
- Is it possible to achieve sub 2 seconds convergence time using linux 
and quagga/ospf apart from modifying the min_delay? I am sure that I am 
not the first person that try to have this kind of performance ...

I don't think that multipath routing is causing the issue but I can test 
without it, of course.

Thanks
Francesco






More information about the Quagga-dev mailing list