caitlinb at broadcom.com
Fri Nov 3 09:07:25 PST 2006
Larry Kreeger (kreeger) wrote:
> I was catching up on old RBridge threads and saw this one about LIDs.
> LIDs can be useful in reducing costs in large switches where
> the entire (potentially large) local MAC table would not need
> to be present in every module of the system. In fact it
> could probably eliminate the need to keep locally learned
> MACs in hardware at all. Assuming a switch with 16 slots and
> 48 ports per linecard, and servers running 16 virtual
> machines each connected to the ports, the number of local
> MACs would be 16*48*16= 12,288. With 48 bits per address,
> that is ~59kbits.
> With LIDs, an entry is need only per port, not per local MAC.
> So given a 16 bit LID, only 16 cards * 48 ports * 16 bits =
> ~12kbits, a savings of ~47kbits.
> The more locally attached MACs, the bigger the savings.
Instead of keeping the MAC to Port mapping in the egress RBridge
you instead keep it in each ingress rbridge that has learned
of this rbridge. I'm not seeing a net savings in memory costs
anywhere. There is certainly a potential latency benefit, but
is there really one that is valid over the desired lifespan
of the protocol?
More information about the rbridge