Hi!

[Roland Karlsson]

Thx Eddie,

We are currently making a vanilla version that can handle say 10.000
packet/s. But ... we are soon going to make some experiments with drivers
to boost that performance. We are doing some manipulation of the data,
so 100.000 may be out of reach with any drivers. According to 'top'
on Solaris and Linux we use 14% CPU when handling 10.000 packets though.
Don't know whether that information is reliable.

Roland



----- Original Message -----
From: Eddie Kohler <eddietwo at naugabutt.lcs.mit.edu>
To: Roland Karlsson <roland.karlsson at terraplay.com>
Cc: <click at amsterdam.lcs.mit.edu>; Magnus Jandel <Magnus.Jandel at terraplay.com>
Sent: Monday, February 12, 2001 9:49 PM
Subject: Re: Hi!


> Hi Roland et al, sorry for the delay in replying!
>
> > So, how would you do to get at least say 50,000-100,000 packets?
> >
> > Do we need to run the router at kernel level?
> >
> > Do we need to implement a polling driver?
>
> You need to implement a polling driver and the architecture around it. Then
> you might be able to run yourrouter even at user level. (We haven't tried
> this case.)
>
> It depends though., on the architecture and many other things. With x86
> Linux machines, kernel-level routers, and no polling, we could get up to
> around 100,000 packets per second. The code we used to get this is
> distributed with Click, in the tools/udpgen directory.
>
> But polling is required for faster speeds.
>
> > Will we benefit from interfacing to the Click system?
>
> Why, sure! Particularly if you use Linux on x86, and have access to Tulip
> cards.
>
> Please let us know how your experiments go.
>
> love,
> ed
>