Re: [lwip-users] performance

[Sam Jansen]

On Thu, 2004-07-29 at 21:50, K.J. Mansley wrote:
> On Wed, 2004-07-28 at 22:53, Sam Jansen wrote:
> > Where lwIP doesn't shine is when the network coniditions are not so 
> > perfect. It's my belief that the granularity of the timers hurts it 
> > here, having to wait a minimum 500ms for a retransmit timeout can be 
> > seriously detrimental to TCP performance.
> 
> That is, I think, something which is common to all (most?) TCPs, not
> just lwIP.  It's also something that would be easy to change.  I agree
> that, if TCP was being specified now, it would be unlikely that the
> retransmit timer would have such a long timeout.

Maybe in the 1980's.

If we look at FreeBSD as a shining example of a BSD TCP implementation,
you'll find that the timer granularity by default is 10ms. This is set
by the HZ value which is specified in the kernel configuration. With a
HZ of 100 (the default), the softclock() function is called 100 times a
second, which will check the timers.

Often people compile with HZ set higher, recent values I've heard quoted
are 1000 and 4000.

Also, to help RTT estimation most operating systems support the TCP
timestamp option.

> Apart from the changes that are in progress at the moment, thanks to
> suggestions from you and Karl Jeacle, is there something we're not doing
> that other stacks are, or is it just that our implementation of the
> algorithms is less good?

I'm still trying to answer this question myself. My current thoughts are
that the Newreno congestion control algorithm is missing, that along
with finer grained timers and better RTT estimation might help a lot.

> It seems you're able to compare against a number of other stacks, which
> is something we've not done a great deal in the past, and is obviously a
> good idea.  If you're able to make any other suggestions of specific
> areas for improvement, I'm sure we'd be interested to hear them.

No doubt there will be more input from me in the future. My PhD research
involves comparing FreeBSD 5.2, OpenBSD 3.5, Linux 2.4 and lwIP in a
simulation context at the moment, so I have quite a good testing
framework.

-- 
Sam Jansen                                           address@hidden
Wand Network Research Group             http://www.wand.net.nz/~stj2