frank@CAEN.ENGIN.UMICH.EDU (Randy Frank) (04/10/90)
I've sat quietly on the sidelines listening to various Ethernet nonsense until
I can't take anymore :-)
1) the reason why you typically see 1 to 2 megabits per second process-to-
process typically has NOTHING to do with Ethernet loading. It has almost
everything to do with:
a) poor ethernet controllers that can't sustain more than a couple of
megabit/second of sustained transmission
b) poor implementation of TCP. Prior to the work of Van Jaconsen virtually
no ones TCP got more than a couple of megabits/second process to process.
c) layering effects of the protocol implemetations, process context switch
times between different protocol layer implementations, and (often
significantly) buffer copying times within a host.
d) lots of other implementation issues within a host that have the effect
of slowing down actual process-to-process bandwidth
e) gateways (routers) in the transmission path that often severely limit
the effective transmission. Gateways often become severly overloaded
well before the subnets to which they are attached.
2) Ethernets under load
a) Ethernets under load often to reach 8 to 9 megabits/second of useful
carrying capacity.
b) When load increases while collision do increase it is controlled.
Remember, just because two hosts want to transmit DOES NOT mean that
a collision occurs. Ethernets run under carrier sense W/ collision
detection: translation: listen before you transmit, and only transmit
if you hear a clear channel; backoff if you hear a collision while
transmitting. There is a very small window (50 usec) under which a
collision can occur. Two hosts have to try to initiate a transmission
within 50 usec of each other in order for a collision to occur; in all
other cases they will notice an active carrier and wait.
c) in the event that a collision does occur, it will be noticed within the
same 50 usec window, and both hosts backoff immediately. On a heavily
loaded ethernet the most likely scenario is that a third host will grab
the wire immediately. (no one said that Ethernets exhibited guaranteed
fairness).
3) the effect of point 1 above is that it is VERY difficult for a single host
to in fact use up anywhere close to the bandwidth of the network.
Randy Frank