RE: Ethernet Characteristics re: packet loss
> The sustained average network throughput you quoted depend on the speed of
> the CPU. A dual PIII running at 1
> GHz or higher (with a FSB running at 133 Mhz or higher) can achieve 80-90%
> utilization of one gigabit link. Nevertheless once you start to aggregate
> multi-gigabit links together you hit a ceiling that can not be overcome by
> additional CPU cycles and TCP/IP Offload becomes a necessity. Hence TOE is
> not only a must for 10 Gb Ethernet but also a requirement for
> gigabit link aggregation.
You are quite correct. The number I was quoting were off a 700 MHz dual
PIII. Actually, it was a Dell :-) The number also varies with the adapter
used. Of course, you could also get higher percentage by moving larger
files with a small percentage of control messages. Thank you for pointing
out the need of TOE even for Gigabit link.
Antal Gyori wrote:
> I am an EE student at Caltech and I am following the ethernet mailings
> because I am interested in networking. I wanted to ask if you know where I
> could find more information on TCP Offload Engine (if there are any of
> them out there, if people are working on them, other references).
Almost every company in business of Internet hardware has something to do
with TOE these days. Please contact Costa at CISCO who is leading a WARP
effort that is a part of the iSCSI WG defining a Remote DMA function needed
by all TOE adapters. You might even wish to get on the WARP reflector. I
copied Costa on this email.