Re: [PATCH 2/2] ipv4 routing: multipath with cache support, 2.6.10-rc3

[Einar Lück <lkml@xxxxxxxxxxxxxx>]

David S. Miller wrote:

> So essentially you want per-flow multipathing.  Except that you're 
> implementation
> is over-optimizing it to the point where it's only per-flow for your specific
> case where the connections are short lived and high rate.
>
> This hurts long lasting connections.
>
> So I'm pretty much against this change.  Do it right by making it occur
> per-connection attempt, it's not my problem to figure out how to do that
> efficiently, it's your's :-)

We do not want per-flow multipathing. We want per connection multipathing with 
fast route lookups (that's why we have all routes in the cache). That is 
exactly what we implemented. Our tests prove that a connection keeps its route 
as long as it lives (the dstentry remains associated with the socket). That's 
why I do not really
understand why our approach hurts long lasting connections in any way. Can you 
explain your point more precisely?

Regards,
Einar.