In article < XXXX@XXXXX.COM >,
: Today we have a T3 link that connects two buildings together. We are
:currently using an MTU of 4352 for this particular link. I am considering
:replacing this WAN T3 link with a true 100Mbps Ethernet link (Which would
:now set my MTU size down to ~1500) to connect the same two sites. Currently
:our T3 link is being utilized no more the 45% at any one time. The benefits
:of having a higher MTU, as I understand it, is that more stations could talk
:at once... or better put, instead of everyone standing in one long line to
:make it across the WAN link, there are multiple lines (I guess 4:1 or 3:1)
:that can be used.
:- Am I describing the benefits of having a higher MTU correctly?
Not even close.
When there is a large burst of data to send (e.g., ftp, but not
telnet or NFS version 1) then having a large MTU allows more of that
data to be sent at one time. In some cases, that can make a big difference
for performance, as you eliminate the packet-header overhead and
intra-packet gap times and reduce fragmenting.
If there does not happen to be large bursts of data to send, then
the longer MTU makes no difference.
When a large burst of data is sent by virtue of the larger MTU, then
everyone else who is trying to send data has to wait longer for
their turn, because it takes longer to send the larger packet over the link.
Especially if the large packets tend to come from one particular source
and not the other sources, that -increases- latency for everyone else.
Whether you should expect lower latency or not when you go to
an MTU of 1500 is not a simple question. It depends on how your
link is being used. For example, if those large packets are not
sent often, but there a lot of smaller packets, then sending that
large packet in a burst might then allow that particular sender
to go quiet, leaving less contention for the line, and less collisions.
With the smaller MTU, then that large packet would have to be
fragmented, potentially leading to increased collisions as that
sender tries to gain primacy on the link. But the timings are
crucial to the behaviour -- if you research the topic of the
"ethernet capture effect" you will see some hints of how the
exact packet timings can make a big difference in how different
systems are granted access to the line.
You don't happen to mention whether the new line will be full duplex or
half duplex. Switches and a full duplex link should result in orderly
queuing and buffering rather than in the kind of contention scenarios I
describe in the previous paragraph. The capture effect occurs only
with shared collision domains, which doesn't happen if you are fully
switched, every NIC going directly to a switch port. The capture effect
-can- occur if you have hubs, or if you have a shared media such as
thinwire instead of Cat5.
Take care in opening this message: My grasp on reality may have shaken
loose during transmission!