[lwip-users] Re: [lwip] lwIP on Gigabit Ethernet

[Chris Borrelli]

Leon,

Thanks for your response.  We have done the bare hardware + bare software
test.  We get about 40 Mb/sec.  So that is our base.  I'll try out these other
tests as well.

> Also, how often do you call the TCP fast timer routine?

I call the fast timer routine every 200ms.  I have been suspicious of this
from the start.  I call the slow_tmr() function every 500ms.  It seems like
the slow_tmr function is when lwIP goes through the active pcb list and
empties the unsent queues.  Is that correct? or am I totally wrong..


 Thanks,
 Chris




Leon Woestenberg wrote:

> Hello Chris,
>
> in such a case I would start some profile testing:
> 1) start with the Ethernet device driver sitting in a while(1) loop
> sending out Ethernet packets.
>
> This is the maximum throughput o/t "bare hardware + bare software".
>
> 2) Do the same for IP packets (checksum overhead etc.)
> 3) Try the same for TCP packets (maybe you can mimick an extremely big
> receive window
> on the receiving side?)
>
> Also, how often do you call the TCP fast timer routine?
>
> I do not have "I told you so" remarks. The source code is there.
>
> Regards, Leon Woestenberg.
>
> ----- Original Message -----
> From: "Chris Borrelli" <address@hidden>
> To: <address@hidden>
> Sent: Wednesday, June 26, 2002 11:54 PM
> Subject: [lwip] lwIP on Gigabit Ethernet
>
> > Just putting some information out there..
> >
> > I have a simple lwIP (0.5.3) echo application running on a gigabit
> > Ethernet network - no OS, no sockets API, just the raw lwIP API.  The
> > performance is much less than I thought it would be (~4Mb/sec).  Does
> > anyone have any comments/suggestions or "I told you so" about this?
> >
> > The system that its running on is PPC405 @ 200Mhz with an on chip
> > Ethernet MAC (one of our own).  The bare metal performance of the system
> > (no tcp stack) is about 40Mb/sec - the MAC host interface and driver I
> > am working with is not at all optimized (16-bit interface @ 50Mhz, no
> > DMA, no interrupts...).  But adding lwIP drops the performance by 10x
> > (from 40 downto 4).  That doesn't seem right...
> >
> > I took a look at the TCP stream.  It looks like host PC fills the window
> > (1024) of the lwIP device, and then the connection is reduced to sending
> > very very small tcp segments back and forth.
> >
> > Has anyone else seen similar issues?  I get similar performance on a
> > 100Mb/sec line (~4Kb/sec data rate).
> >
> > Does anyone have any lwIP performance numbers for their applications???
> >
> >  -Chris
> >
> > [This message was sent through the lwip discussion list.]
> >
>
> [This message was sent through the lwip discussion list.]

[This message was sent through the lwip discussion list.]