Re: (diet-)FIB alternative fib_hlist.c

[Andre Tomt <andre@xxxxxxxx>]

Robert Olsson wrote:
> Hello!
>
> fib_hlist is the smallest and simpliest routing algo we could think of
> it's just a sorted (h)list.
>
> routing (FIB lookup) performance. dst hash is not used.
>
>     fib_hlist   fib_hash  test        routing table size 
>     -----------------------------------------------------
>     444 kpps    433 kpps  Single flow. local=19/main=5 entries
>     433 kpps    431 kpps  rDoS.        local=19/main=5
>     0.2 kpps    198 kpps  rDoS         local=19/main=123946
>
> As seen fib_hlist is catastrophe for large routing tables as expected but 
> performs surprisingly well for ordinary routing tables so it should be 
> fine for most hosts and servers. The patch has config option to select FIB.
> Probably we soon want to specify differnt lookup schemes for different 
> tables say for local table fib_hash or fib_hlist. While for large main table
> fib_hash2/fib_trie would be better option.
From a distribution[1] point of view, a boot time option would be much 
better than choosing at compile time. Boot time plus a kconfig multiple 
selection list would be even better. Another perhaps even more user 
friendly option would be setting it runtime using f.ex. iproute, but I 
guess that could end up messy..
1. goes for both "a GNU/Linux distribution" and kernel image
   distribution in a production network