"I know there are a lot of other factors (like packet switching,
collisions, etc)"
Collisions wouldn't be a factor, in this case, due to the machines in
question running at 100Mbps/full duplex. One item I haven't seen
mentioned is the CPU utilization on the machine with a weak processor.
Have you tried running "top" during a large transfer to see if the
CPU is able to keep up?
Kurt's statements about protocol overhead are accurate. You'll never
get anything approaching the theoretical throughput. One of the
tricks a former employer of mine used to like to talk about was 20Mbps
(it was a long time ago) that you could get through their system,
because it was 10Mbps/FULL duplex. So theoretically, because the
information can flow in both directions at 10Mbps, you *could* get
20Mbps, if you were transferring information in both directions, at
full speed. Of course, that's the theoretical limit, and it isn't a
situation that occurs very often in the real world, anyhow. But,
marketing people like big numbers...
On 4/17/05, Jim <
jnantz2@216-19-216-108.getnet.net> wrote:
> Thank you for the information. One of these days I'll get ambitious and
> buy a couple of gigabit cards.
>
> Kurt Granroth wrote:
> > Not much more, in any case. The general rule of thumb for a 100Mbps
> > network is 40% efficiency.
> >
> > My knowledge of this is sporadic, rusty, and incomplete... but here goes:
> >
> > 100BaseT dictates that each 8 bit byte is encoded in a 10 bit packet.
> > Therefore, the theoretical max throughput of raw bits would be about
> > 76Mbps (megabit would be 1024x1204, not 1000x1000 when talking about data).
> >
> > Nobody pushes raw bits over the network, though. We use Ethernet on one
> > layer and then IP and then TCP or UDP. All of these layers have their
> > overhead. Let's take TCP/IP over Ethernet. A maximum Ethernet packet
> > is 1514 bytes. Of that, 54 or 66 bytes (depending on data type) are
> > taken by the Ethernet, IP, and TCP headers. That's about 4%. But then,
> > TCP requires an ACK packet for each and every data packet. That means
> > that the headers actually take up about 8% of the space. Subtracting
> > that from our theoretical max shows that the TCP/IP over Ethernet has a
> > theoretical max of 70Mbps.
> >
> > Remember, though, that up to now, we are talking about theoretical
> > maximums. The real world is quite a bit more variable. Here's where my
> > knowledge gets increasingly iffy. Take each statement from this point
> > on with a big grain of salt.
> >
> > Okay, so the theoretical max is 70Mbps but that assumes that each packet
> > is maxed out with data. If we go to a worst case scenario where each
> > data packet only contains 1 byte, then the actual throughput would be
> > less than 1Mbps. Say the packets are each (on average) half full.
> > Then, our actual throughput would be under 60Mbps.
> >
> > Beyond this, I admit that it gets into voodoo territory. I know there
> > are a lot of other factors (like packet switching, collisions, etc) that
> > work to drive the actual throughput lower and lower but I don't know
> > enough about those factors to really comment.
> >
> > The end result, though, is that 40% is the highest you can reasonably
> > expect. So 33Mbps is getting pretty close. Maybe you could eek another
> > 4 or 5Mbps out of that that with some careful tweaking... but that's
> > going to be it.
> >
>
> ---------------------------------------------------
> PLUG-discuss mailing list - PLUG-discuss@lists.plug.phoenix.az.us
> To subscribe, unsubscribe, or to change you mail settings:
> http://lists.PLUG.phoenix.az.us/mailman/listinfo/plug-discuss
>
---------------------------------------------------
PLUG-discuss mailing list -
PLUG-discuss@lists.plug.phoenix.az.us
To subscribe, unsubscribe, or to change you mail settings:
http://lists.PLUG.phoenix.az.us/mailman/listinfo/plug-discuss
(text/plain)