two GbE interfaces and maximum bandwidth achievable

Hello all,

I have been posed the following question: "A PC motherboard has two GbE
interfaces, and is acting as a bridge. The theoretical maximum bandwidth
achievable across the bridge is 2Gbps. In practice only about one half of
this is achieved. Why might this be?"

First of all, I cant really understand if the term bridge means layer 2
switching or its just a misuse of the term and means in reality layer 3
routing. Can you give me some pointers regarding the reasons why only half
the theoritical bw can be achieved?

(in case bridge means in reality layer 3 routing, inefficient scheduling on
a perhaps loaded systems may slow down the routing process).

TIA

"jimjim" <(E-Mail Removed)> wrote in message
news:RtOKe.87020$(E-Mail Removed). uk...

> I have been posed the following question: "A PC motherboard has two GbE
> interfaces, and is acting as a bridge. The theoretical maximum bandwidth
> achievable across the bridge is 2Gbps. In practice only about one half of
> this is achieved. Why might this be?"

It all depends how the GbE interface connect to the CPU.

> First of all, I cant really understand if the term bridge means layer 2
> switching or its just a misuse of the term and means in reality layer 3
> routing. Can you give me some pointers regarding the reasons why only
> half the theoritical bw can be achieved?

I'm presuming the person asking the question knows they mean to ask and
is talking about bridging, not switching.

> (in case bridge means in reality layer 3 routing, inefficient scheduling
> on a perhaps loaded systems may slow down the routing process).

It could also slow down the bridging process.

The problem is usually either the bandwidth to the GbE interfaces or the
efficiency of interrupt handling. Ideally, you want the interfaces to buffer
many packets, issue a single interrupt, and all the CPU to tightly loop
through its bridging code queuing the packets to the other interface without
the other interface interrupting it each time it sends one of the packets.

DS

"jimjim" <(E-Mail Removed)> écrivait
news:RtOKe.87020$(E-Mail Removed). uk:

> Hello all,
>
> I have been posed the following question: "A PC motherboard has two
> GbE interfaces, and is acting as a bridge. The theoretical maximum
> bandwidth achievable across the bridge is 2Gbps. In practice only
> about one half of this is achieved. Why might this be?"
>
> First of all, I cant really understand if the term bridge means layer
> 2 switching or its just a misuse of the term and means in reality
> layer 3 routing. Can you give me some pointers regarding the reasons
> why only half the theoritical bw can be achieved?

BRIGING or ROUTING with two "inline" GB interfaces give you 1GB in each
direction, not 2GB.

TRUNKING with two "paralelle" GB interfaces give you 2GB in both
direction. You must use the same trunking protocol between the two
components (pc<->switch with Cisco trunking, or linuxbox<->linuxbox with
linux trunking).

> (in case bridge means in reality layer 3 routing, inefficient
> scheduling on a perhaps loaded systems may slow down the routing
> process).

In any case, harware architecture (bus efficiency) and cpu load may
reduce the bandwidth.

On Linux, there is a fast path network option (at compile time if I
remember).


Regards

> It all depends how the GbE interface connect to the CPU.
>
Do you know any possible ways?

> I'm presuming the person asking the question knows they mean to ask and
> is talking about bridging, not switching.
>
I thought a switch is a multi-port bridge though! Do they have any
fundamental functional differences?

> It could also slow down the bridging process.
>
Is there a possibility that the bridging process is carried out by hardware,
just as in modern switches?

> The problem is usually either the bandwidth to the GbE interfaces
can you elaborate, plz?

> or the efficiency of interrupt handling.
true! fine-grained clock interrupt rate.

TIA

> BRIGING or ROUTING with two "inline" GB interfaces give you 1GB in each
> direction, not 2GB.
>
Can you please define the term inline for me? and then elaborate a little on
your point? Do you mean that the vendon over-specified his mobo?

> TRUNKING
this is used to carry traffic from multiple VLANs though, isnt iit?

> In any case, harware architecture (bus efficiency) and cpu load may
> reduce the bandwidth.
>
hmm, true. I am afraid though this was not the answer he was waiting for.

TIA

"jimjim" <(E-Mail Removed)> wrote in message
news:Z58Le.87766$(E-Mail Removed). uk...

>> It all depends how the GbE interface connect to the CPU.

> Do you know any possible ways?

PCI, PCI-E, and Intel's direct to northbridge thing, I think they call
it CSA.

>> I'm presuming the person asking the question knows they mean to ask
>> and is talking about bridging, not switching.

> I thought a switch is a multi-port bridge though! Do they have any
> fundamental functional differences?

I'm sorry, I mean to say "bridging, not *routing*".

>> It could also slow down the bridging process.

> Is there a possibility that the bridging process is carried out by
> hardware, just as in modern switches?

Not if you're talking about a commodity PC with two GbE interfaces.
There is nothing other than the CPU to do it in that case.

>> The problem is usually either the bandwidth to the GbE interfaces
> can you elaborate, plz?

If you put two GbE interfaces on a 32-bit 33Mhz PCI bus, and each packet
has to traverse the bus twice (once to be read in and once to be written
back out), the maximum bandwidth is 500Mbps total in both directions
(130MB/s is about 1Gbps, and you halve that). That means you could, at best
with perfect efficiency, handle 1/4 the possible traffic.

>> or the efficiency of interrupt handling.
> true! fine-grained clock interrupt rate.

You want one interrupt to occur and then the CPU to read multiple
packets without the outbound interface interrupting it after it sends each
packet. This is usually called "interrupt coalescing".

DS

jimjim wrote:
>>BRIGING or ROUTING with two "inline" GB interfaces give you 1GB in each
>>direction, not 2GB.
>>
>
> Can you please define the term inline for me? and then elaborate a little on
> your point? Do you mean that the vendon over-specified his mobo?
>
>

I think his point is that if you transfer a file, or test data in a
single direction the fastest bitrate you will see is 1Gb/s. 2Gb/s comes
when you have data going in both directions at the same time.


>>TRUNKING
>
> this is used to carry traffic from multiple VLANs though, isnt iit?
>
>
>>In any case, harware architecture (bus efficiency) and cpu load may
>>reduce the bandwidth.
>>
>
> hmm, true. I am afraid though this was not the answer he was waiting for.
>
>

Check you drivers too. Nvidia drivers have a setting that can be
throughput, or CPU. Throughput maximizes throughput at the expence of
heavy load on the CPU which can make other things like sound, etc.
choppy. Probably a good setting for a server. CPU tries to keep things
smooth to the users appearance.



--
Respectfully,


CL Gilbert

"CL (dnoyeB) Gilbert" <(E-Mail Removed)> écrivait
news:-YGdnbkeAYrJn2PfRVn-(E-Mail Removed):
> jimjim wrote:
>>>BRIGING or ROUTING with two "inline" GB interfaces give you 1GB in
>>>each direction, not 2GB.
>>>
>>
>> Can you please define the term inline for me? and then elaborate a
>> little on your point? Do you mean that the vendon over-specified his
>> mobo?
>>
>>
>
> I think his point is that if you transfer a file, or test data in a
> single direction the fastest bitrate you will see is 1Gb/s. 2Gb/s
> comes when you have data going in both directions at the same time.

That's it.

In this case the data pass throw only one nic at a time and in one
direction only.

>>>TRUNKING
>>
>> this is used to carry traffic from multiple VLANs though, isnt iit?

Trunking is a way to agragate ethernet trafic. Normally it's used between
two cascaded switches to avoid congestion on the backbone.

But if you have many nic in a server you may use trunking to agragate
them between the serveur and the switch.

You need to setup the same trunking protocol on the server and on the
switch (Cisco, 3Com, ...).

In that case, the switch (and the server) will use the two patch, and you
will have 2GB/s in each direction.


Regards