[lwip-users] Re: [lwip] tcp flushing

[Kieran Mansley]

On Fri, 29 Nov 2002, Tim Newsham wrote:

> > I think it should just be documented so people can optimize for either
> > latency or throughput. It can even change inside the same code. For example
> > a telnet server should be low latency, while ftp transfers should be high
> > throughput.
> > I don't vote for the socket interface, only the raw interface (I think this
> > is one of the reasons it's called raw).
>
> I agree with the documentation bit, but I disagree that the raw interface
> should not support automatic flushing.  I think its a bug that a tcp_write
> would NEVER get flushed without implicitely calling tcp_output or having
> an external event occur.  Leaving this behavior in tract is gonna bite
> someone else at some point.  Should be fairly easy to have something in
> slowtmr call tcp_output on any pcb's that havent been flushed in a while,
> no?

In my opinion, flushing in slowtmr is worse than never.  If it happens
then, it could be delayed by half a second, but the fact that it happens
may mean the person using lwip thinks that the poor performance is just a
property of lwip, rather than realising that they're not using the
interface as it was intended.

Kieran

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