On Sat, 2004-01-24 at 14:52, Leonid Grossman wrote:
> > What would also be interesting to see is a packet that never
> > leaves the kernel such is in forwarding. If theres a way you
> > can stash two of those cards in a box and just have them
> > forward packets coming in from one to another - would be nice
> > to see the numbers with varying packet size example {
> > 64,256,512,1518, 4K, 9K}
> >
> > cheers,
> > jamal
>
>
> Hi Jamal,
> Are you suggesting to run a benchmark between two back-to-back hosts,
> and then run the same benchmark between the two hosts via a third box
> (with two 10GbE cards) that would just forward traffic, and compare the
> results?
> Or you had a different setup in mind?
Just a simple test for packets that dont cross user space - kernel
boundary. Example forwarding of some form where a packet comes in,
data gets touched (ttl, csum etc) and gets forwarded to the egress port.
The setup will include a single box with two NICs; one connected to a
traffic generator source the other to a sink which records stats.
The interesting measurements will packet latency and throughput.
Get a off the shelf traffic generator like an IXIA (pktgen may be used
too); i.e forget ttcp and relatives.
For starters you could even have a packet coming in and being DMAed to
the otehr NIC untouched.
> Not sure what the application for something like this would be (I think
> 10GbE will be mainly deployed in a datacenter for a while), but we can
> probably run something like that and get the numbers; I'll let you know.
>
Apps would be any middle box (router, firewall, accounting etc).
Interest is more from a linux side what can we do to improve things.
cheers,
jamal
|