Not every bit in results in just one bit out. Broadcast, multicast,
flooding for unknown MACs (or switching failures), ...
They were talking about a simple scenario where a bit that enters a port
will leave a port. With 24 gigabit ports, for all intents and purposes,
you will only ever have 24 gigabits at the most traversing the backplane.
Doesn't that assume that the communicarion is unidirectional?
If two hosts are exchanging 1Gbps flows, the traffic across the bus will be 2Gbps, right?
And of course, this doesn't include any bus-intensive operations like multicast
or things which require cpu processing - those can consume a lot more resources than the input rate of the port.
-David Barak
Tom Storey wrote:
Were not considering anything other than basic switching in this scenario, as is my understanding.
2 hosts will create 2gbps of traffic as each host is inputting 1gbps into the switch (just multiply it by 12 to give you 24 ports). 3 hosts will create 3gbps of traffic as each inputs 1gbps into the switch (e.g. each host could be sending 500mbps to each of the other hosts). And thus and so forth. 
You can only input a maximum of 24gbps into the switch, which means that only 24gbps will cross the backplane.
Yes there is 48gbps if you combine tx and rx of each port, but traffic only has to cross the backplane once, from rx on one port to tx on another.
Sorry if I have hijacked this thread from the OP.
Tom
Doesn't that assume that the communicarion is unidirectional?
...
No.
If two hosts are exchanging 1Gbps flows, the traffic across the bus will be 2Gbps, right?
Yes. 1Gbps backplane impact per host. You have two hosts, right? One host per port? That's 1Gbps per port.
So, 24 ports = 24Gbps, right?
Let's try look at it another way:
- A 24 port gig switch can receive at most 24Gbps.
- That same switch can transmit at most 24Gbps.
You don't get to add transmit and receive together to get 48Gbps. Packets don't go across the backplane once to receive, and then once more to transmit. They go across once, from the receiving port to the transmitting port. (sure, sometimes perhaps packets do go across twice, but not normally)
And of course, this doesn't include any bus-intensive operations like multicast
or things which require cpu processing - those can consume a lot more resources than the input rate of the port.
Of course multicast/broadcast consumes more resources than the input rate. That's the point. If you receive multicast or broadcast at 1Gbps, and the multicast needs to go out all the ports, you need to transmit at 24Gbps. That's 24 x the transmit resources (and probably backplane resources, depending on architecture etc. etc.) than a single 1Gbps unicast stream.
Of course, with unicast it is only getting to one host.
Let's assume we have data at 1Gbps that we need to get to 24 hosts.
- If we unicast, we need 24 input ports, and 24 output ports, assuming we only have gig ports (or say 3x10GE, or whatever).
- If we multicast, we need 1 input port, and 24 output ports.
When you compare the end result, multicast uses significantly less resources, right?
In fact, perhaps some bus architectures know about how multicast works, and it consumes *less* resources than doing the same thing with many unicast streams. If the bus does not know about multicast, then the bus would treat it as 24 unicast streams, surely.
Nathan Ward wrote:
If two hosts are exchanging 1Gbps flows, the traffic across the bus
will be 2Gbps, right?You don't get to add transmit and receive together to get 48Gbps.
Packets don't go across the backplane once to receive, and then once
more to transmit. They go across once, from the receiving port to the
transmitting port. (sure, sometimes perhaps packets do go across
twice, but not normally)
Assuming a crossbar switch, sure. If your ports individually look up
the outgoing port for an incoming packet, request backplane to that
port, and transmit, then you only need 24Gbps. If your ports need to
connect to an intelligent entity on the backplane to do your
routing/switching/IGMP snooping/QoS enforcement/etc, then you are indeed
going to cross the backplane twice, and need both transmit and receive
bandwidth.
Since many of us are routing goons with store-and-forward roots, we tend
to think along those lines. And it is still wise, even in this day and
age, to make sure that backplane bandwidth doesn't include a central
switching point, or, if it doesn't, the marketing folks haven't doubled
the backplane numbers because they took it out.
-Dave