Re: Re: Re: [lwip-users] Re: tx badnwidth

[Kieran Mansley]

On Thu, 2009-03-05 at 10:47 -0500, Chen wrote:
> Thanks for your suggestion
> 
> I increased the TCP_SND_BUF to 32000 and PBUF_POOL_SIZE to 16, and the
> tx and rx ends are fighting now, trying to resolve the tcp window
> update, it seems
> 
> please see http://www.dataq.com/chen/new.cap

I don't see any fighting there: can you explain what you mean.  It looks
like a packet was lost, lwIP sent duplicate ACKs, the other end resent
it, lwIP acked the retransmitted packet, and they're both now waiting
for the other to send some data.  Which end should have data to send at
this point?  They're both got available window to send if they want to. 

> What is the rule of thumb to set up these two parameters? Can the
> first be be larger than 32K?

Yes, I think so.  They're probably restricted in lwIP by the size of the
type used to store them: take a look at the source.  There isn't really
a useful rule of thumb: larger will mean higher throughput until
something else becomes the bottleneck, then setting them any bigger will
just waste memory.  The best value depends on your system's needs to
trade off performance vs memory, the traffic pattern you expect, the way
the application uses the API.

By setting PBUF_POOL_SIZE to 16, and with each pbuf at 500 bytes (as
specified in your lwipopts.h) you've got a maximum of 7500 bytes of
buffering for all the users of pool pbufs.  In reality not every pbuf
will be completely filled, so you'll have a bit less than that
available.  Therefore setting TCP_SND_BUF to 32000 is probably not going
to make much difference compared to setting it to 7500, as you could
never put 32000 bytes in the send buffer - there aren't the pbufs
available to do it.

Kieran