Re: Very poor TCP/SACK performance

Mark Gray (markgray@iago.nac.net)
11 Sep 1998 12:56:16 -0400


alan@lxorguk.ukuu.org.uk (Alan Cox) writes:

>
> > Maybe Linux should reduce the window size if it detects a
> > lot of out-of-order packets coming in. If the TCP code has
> > a way of knowing the properties of the connection networking
> > is done over, it might be able to set these flags itself.
> >
> > ie: if(small_buffer_other_side && lots_out_of_order()) size--;
> >
> > Networking guys, what about this?
>
> The solution you describe required an additional telepathic link
> between the two modems so you can tell queue drops from lost frames
> it also assumes the only bottle neck is the last hop and that the transit
> path has a low bandwidth*delay product.
>
> Alan

An interesting approach to custom configuring the TCP/IP - ppp
algorithm might be to write a userland program which would analyse a
tcpdump of an interface's traffic and suggest parameters you could
then cat to /proc/* the next time the interface was up (or possibly on
the fly even.) To carry the overkill a step further you could run the
tcpdumps collected over time through another program which would
evaluate the first program's algorithm and suggest ways to fine tune
it. (Designing it all so that your learning system didn't decide that
unfettered greed was the best policy is left as an exercise for the
reader. :-) For linux-9.1 you could hack named to cache configuration
information as a psuedo-record type for every destination and then
have custom configurations for each and every TCP address which has
seen statistically significant traffic.

Volunteers?

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/faq.html