On Sun, Sep 22, 2013 at 9:56 PM, Kurt Buff <kurt.b...@gmail.com> wrote: > Well, I do remember reading a long time ago that traffic shouldn't go > through more than three switches on a LAN (was that referred to as the > diameter? I can't remember) - that pretty much matches the Cisco model > of core, distribution and access, as described here, among many other > places:
Well, there's a difference between "not a best practice" and "violates the specification". The 5-4-3 rule arises because of how classic Ethernet works. It's a shared medium, meaning all nodes on the network are connected to a common signalling bus. It takes time for a signal to travel through the bus. If the bus is too long (or has too many repeaters), a signal transmitted from one end may not reach the other end before various time windows close, leading to things like missed collisions, missed preambles, and the like. Worse, they won't be retried by the link layer when they should be. (Ethernet does not guarantee delivery, but it does have *some* retry logic built-in.) Silently corrupted payload is even possible. (The "5-4-3 rule" is technically only a guideline because Ethernet doesn't care about repeater hops, it cares about signal timings. The only way to truly confirm adherence to all aspects of the specification was with a network analyzer. Those are expensive, so a rule of thumb was needed. 5-4-3 worked for most everything. If your equipment happened to need less margin within the spec, your network could be bigger.) Modern Ethernet puts a transceiver at each end of each cable. Each cable is a point-to-point link, as far as the MAC sublayer is concerned. The switches receive and buffer frames, as opposed to simply amplifying the electrical signal, the way a repeater does. Assuming full duplex, you don't have collisions at all, so you don't have to worry about jam signal propagation. You don't have to worry about preamble degradation, since each switch is transmitting a new preamble. However, if you have to go through 42 switches to get to your destination, that's still bad. Ethernet still does not guarantee delivery, and each link is another chance for a frame to be corrupted and discarded. Each switch also introduces more latency, and a lot of LAN protocols (SMB, I'm looking at you) *hate* latency. -- Ben