Re: [lwip-users] Performance Tuning?

[David Ryan]

My performance checking has been with the latest version of lwip.  I've 
written a small java app and server that simply receives the data via a 
socket connection.  No action is taken on the data, so its basically the 
equivalent of a net /dev/null.   Sending data to the ps2 is the most 
common use of the stack so this seemed to be a reasonable test.  (I also 
did some round trip testing using an echo server).

I also used ethereal to capture the communications so I could see what 
was going on.  What I found was that the TCP_WND(32Kb) was reducing to 
zero in about 50ms.  It was then taking another 50ms for the data to be 
removed from the stack before an ACK was being sent with the TCP_WND 
back at 32Kb.

I am yet to work out why the server thread is not reading the data off 
the stack  (could be related to thread priorties ).   I'm assuming if I 
can fix this problem I should be able to double the performance.

David.

> Hi David,
>
> I would really recommend you port your code to the latest version of 
> lwip. It has many bug fixes and performance improvements.
>
> Take a look at the coldfire version of lwipopts.h in the 
> contrib/ports/coldfire area. I tried hard to optimize for some 
> performance. I think lwip is quite sensitive to the pbuf size. You 
> want to make sure that is a decent size (1024 or bigger).
>
> One other thing you want to watch out for is how you output data via 
> TCP. You want to make sure your writes have a relatively large amount 
> of data in them. I wrote a telnet server and was using that as a 
> console for my application. I was getting terrible performance, but I 
> found out that the code was basically calling a putc() routine for 
> every character that wound up in a separate tcp frame. No wonder it 
> was so slow. I implemented a buffering scheme in my pseudo-terminal 
> layer and it improved performance immensely. I have'nt measured real 
> performance quantitatively yet. At some point I have to implement an 
> ftp server and I will probably do it then.
>
> How are you measuring performance?
>
> Regards,
> David.
>
>
> David Ryan wrote:
>
> > Hi All..
> >
> > Another question..  I thought I'd keep it seperate from the last post 
> > as its on a rather different subject.  Users of our 
> > implementation(based on lwip 0.5.3) have been complaining that the 
> > performance they get from the stack is around 150kb/sec - 350kb/sec 
> > using a 100mbit ethernet adapter. This seems to be way below the 
> > theoretical maximum and below the 1.5mb/sec people are getting using 
> > ps2linux on the same hardware.
> >
> > I realise that performance can be effected by many things, however 
> > I'd like to know if there is any documentation or experience from 
> > people on how best to tune lwipopts.h for better performance when 
> > dealing with streaming data.  Obviously some values relate to each 
> > other.  The question is which values relate to each other and how..  
> > and how should they be managed to get the best performance.
> >
> > Thanks,
> > David.