scottp@eecs.cs.pdx.edu (Scott Peterson) (08/25/90)
I just read a paper on ethernet switches (as opposed to repeaters, bridges,
and routers) and I have a question. How do these devices handle collisions?
Briefly, an ethernet repeater is a device for connecting multiple segments
which can begin transmitting the forwarded packet to the destination segment
before it has been completely received on the source segment. The theoretical
lower bound on propagation delay is the time from the start of the preamble
to the end of the destination address, plus the prop delay from the source
station to the switch. This is better than store-and-forward because of the
lower prop delay, and better than a monolithic LAN because two packets can
be intransit simultaneously through the same switch as long as all four
stations involved are on different segments (or some of the traffic is segment
local, but that's the trivial case).
If the destination segment is busy, it's too late to signal a collision to
the source station, so the packet must either be dropped, or the switch
must revert to the store-and-forward approach of bridges. Switches already
have to have most of the featurs of a bridge, including the learning
capability. Adding store-and-forward may not be that expensive.
Dropping the packet in the event of a collision is probably a bad idea,
especially in a multi-switch environment where the packet may traverse many
switches (may still be faster than a bridge). In the many-hop case, the
odds of all segments being idle simultaneously is small for a loaded
LAN.
Anyone have experience with these things?
Thanks in advance,
--- Scott Peterson
scottp@eecs.cs.pdx.edu