What levels of oversubscription are common in ethernet fabric backbones in
a LAN setup? Let's say each of my switch ports supports 1 Gbps; does that
mean the switch-to-switch interconnect has to be specified to be "number-
of-switch-ports" Gbps? I suspect not, but what are the common reduction
factors people use?
Also, naively I assume that worst-case I'd get an effective bandwidth of
[ backbone / number-of-switch-ports ]
Is this a reasonable approximation? In this case though, does the switch
communicate to the ports and ask them to "talk" slower than 1Gbps? If it
didn't then this might mean dropped packets since the switch-to-switch
fabric was already full (assuming the switch doesn't have infinite
buffers)?
Finally, how does one apply a similar calculation to predicting the latency
degradation? If a best-case, tiny-packet end-to-end latency is x usecs how
do I predict what it will be after a certain amount of oversubscription in
the switch-to-switch backbone?
--
Rahul
|
Similar Threads
Linear Mode
