Noel,

> However, if the mapping function moves closer to the servers, the cache size
> will scale more reasonably. (A hand-wavy argument that the scaling will be
> feasible is provided by noting that at the limiting case, a single server, it
> can only support a limited number of TCP connections anyway, so it perforce
> will be talking to a limited share of the network.)

I don't think that's even very hand-wavy. If a server can handle state
for N simultaneous transactions, and the extra state per transaction to
cache the mapping info is a small fraction of the total state per
transaction, surely we can pick the cache lifetime so that the cache
size also scales like N.

Also, in the context of large-scale services, "server" these days
means a rack or a room full of virtual servers, which are there
precisely to deal with scaling like N.

So it becomes necessary, not optional, to move the mapping cache
to the places where state has to scale anyway. That takes it
out of traditional networking equipment, I think.

Regards
   Brian Carpenter
_______________________________________________
rrg mailing list
[email protected]
http://www.irtf.org/mailman/listinfo/rrg

Reply via email to