Yeah, jumbo packets are great!  Too bad there
isn't an IEEE standard for them  :-(

> "-ptcpip pagesize=8192,mclbytes=8192"

With enough CPU, you're probably doing gige wire 
rate anyways (assuming you're using PCI-X).

The effect of the above is to defragment the received 
jumbo packets, so that they fit into one buffer (and 
consume one rx descriptor) instead of consuming four 
rx descriptors.  This will reduce your CPU usage and
PCI-X bus usage.

Off the top of my head, going from 2k to 8k MCLBYTES
is going to require four times as much memory for rx.

> if the jumbo packet size is set at 9000

right - you have to decide if exceeding 8K is
really worth it for your configuration.  You
have to look at "pidin mem" for iopkt to see
what memory it's going to cost you, and we had
a version of ttcp around here somewhere that would
display idle cpu percentage during data tranfer
benchmarking.

All of this is about reducing the cpu consumption
of io-pkt during heavy data transfer, by spending
memory.  It's the old size-vs-speed thing again.

--
aboyd