click performance

[Eddie Kohler]

Hi Puneet,

> I have the router running on a machine (P-III 866 Mhz) connected to 3
> 100 Mbps ethernet segments. Since the theoritically maximum packet rate
> that can be generated on a single segment is 148,000 packets per second
> (for 64 byte packets), i can theoritcally generate a load of 444,000
> packets per second on the router and i feel that should be enough to
> exhaust the router's CPU cycles. Is this assumption correct.

It all depends. If you are using polling device drivers (such as the
tulip), then a _700_ MHz PIII, with an optimized IP router, forwarding on 8
ethernet segments, can forward 446 Kpps.

Major factors affecting maximum forwarding rate include:
- interrupting vs. polling device drivers (HUGE)
- PCI bus speed, width, and number (multiple PCI buses can be better,
depending on traffic pattern)
- CPU speed
- routing task
- traffic pattern (if all 3 inputs are going to one output, obviously
you've got a problem)

I would advise you bite the bullet and get some more Ethernet segments. You
might not want the output devices to be the same as the input devices.

> Furthermore,since i plan to use click (udpgen.click)on the machines
> acting as data sources i want to know how many 64 byte packets per
> second can be generated by a single machine (P III 866 Mhz).
> Is this rate enough to saturate one 100 Mbps ethernet segment? Is this
> enough to saturate multiple ethernet segments?

You should have no problem saturating a 100 Mbit Ethernet with a single
machine; we did it with machines much slower than 866 MHz.

> It would be great if there is some information as to how many packets
> can be generated by click
> assuming infinite bandwidth (i.e how many packets per second can be
> generated before the CPU becomes the bottleneck)

That would be great, but it's impossible, since it depends on so many
factors. See our ASPLOS paper for more data, though:

http://www.icir.org/kohler/pubs/click-asplos02.pdf

Eddie