The MPLS WG Archive

[Eric Osborne <eosborne@cisco.com>]

On Tue, Jul 02, 2002 at 10:51:11AM -0400, Jeff Parker wrote:
> > Jeff> There are many functions that map your octets onto the range
> > Jeff> from 1 to 6.  Based on the distribution of your octets 
> > Jeff> (i.e. the flows you expect to see) you might have different
> > Jeff> functions if you tried to spread the load evenly.  
> > Jeff> Wouldn't you need some state to remember which function you 
> > Jeff> picked?
> > 
> > I think the usual implementation technique is to pick a 
> > function and compile
> > it into the code ;-)
> 
> Eric -
> 	There are schemes out there that compile a function
> into code, hash, and then use the result to pick a bucket
> based on ranges computed to equalize the flow.  The 
> state is reduced to N integers, but represents state all
> the same.  
> 	Of course, if you don't care about equalizing the 
> flow, you don't need to be so careful.  


I guess it depends on what you mean by "state".  With the given
definition, one also needs to keep state just to switch IP or MPLS
packets (state about InLabel/OutLabel/OutIF mapping).  Right?  Given
that, any loadsharing "state" is simply an extension of what already
must be kept.  Heck, looking at it this way, even simple, co-like TDM
boxes need state.  I've always looked at state as something computed
dynamically based not on the switching tables but on the traffic
itself (building a flow cache, for example), and let me make it
perfectly clear, I'm not a fan of flow caches[




eric