Claude Morin wrote:
> While the solution you're referring to is for ADSL lines, I think it solves 
> a different problem, where a full-bore upload starves downloads. Besides, 
> Johnny says his upload speed drops to near zero. If memory serves, in the 
> reverse situation (full-bore upload starves downloads), the download speed 
> is still at least 10% or 20% of maximum. Lastly, he says he has 10Mbps 
> up/down (posted after your reply, I think); the starvation issue is most 
> noticeable for asymmetric connections.

Here's a typical account of packet size distribution for
an ADSL line (24 hours of data).  Note the neat bifurcation of TCP traffic --
bulk transfers occur at MTU size,  interactive traffic
and empty ACKs are quite small.

You could start with 2 queues for TCP traffic in each direction,
and base them on packet size, giving greater weight to smaller
packets.

I would simply put the queues as one for packets 0-511 and the other for
packets 511-1500 or whatever your MTU size is.


### Aggregate Packet Size Distribution (including MAC headers) ###
<<<<
  [   32-   63]:      65423
  [   64-  127]:     136835
  [  128-  255]:      28110
  [  256-  511]:       5062
  [  512- 1023]:       4372
  [ 1024- 2047]:      75461

 >>>>>>>>

### ICMP Packet Size Distribution (including MAC headers) ###
<<<<
  [   32-   63]:          4
  [   64-  127]:      18384

 >>>>>>>>

### UDP Packet Size Distribution (including MAC headers) ###
<<<<
  [   32-   63]:          2
  [   64-  127]:      49342
  [  128-  255]:       4690
  [  256-  511]:       1531
  [  512- 1023]:        393

 >>>>>>>>

### TCP Packet Size Distribution (including MAC headers) ###
<<<<
  [   32-   63]:      64921
  [   64-  127]:      67381
  [  128-  255]:      22588
  [  256-  511]:       3350
  [  512- 1023]:       3703
  [ 1024- 2047]:      74795