[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: A question about priorities




Thanks for the suggestions, I have one more doubt which is troubling me.

It seems that traffic control comes into play only when traffic is to be
dequeued to be sent to the output interface. 
Does it also influence how IP layer hands down the  data to the lower layers?

For example, if I have two UDP sockets one set at a higher priority using
setsockopt than the other. Then though the packets for the higher priority
connection are drained first at the interface, if the IP layer hands
equal amount of data from the socket buffers to the layers below (which it will
if Traffic control does not affect how sockets are drained), the observed
throughput for the both the connectons in the long run will be the same,
(since IP hands equal amount of data from both the connections.)

Does this reasoning make sense?

Thanks 
Nikhil

Now suppose that I have 
two UDP connections which take equal turns dumping data whenever there
is place in their sockets.






On Mon, Aug 07, 2000 at 10:37:21PM -0400, jamal wrote:
> 
> 
> I think Andi answered it nicely; but let me add my 2 cents:
> 
> In prio as long as there is data to be dequeued in the higher priority
> queue, prio will favor the higher queue. In the case of the UDP example,
> as you noticed 100% utilization for the higher prio. In the case of TCP, 
> i suspect some packet might have been dropped and TCP backed off and so
> the equilibrium of sharing is reached that way.
> Try, one: 
> 
> - attach a TBF to the higher priority queue and give it bandwidth 10Mbps
> - attach a TBF to the lower prio and give it a b/width of 5Mbps 
> Reverse the above.
> 
> or two:
> - attach bfifo queus to both the low and high priority. 
> Compute the RTT*BW worth of bytes and make that the queue limit of the
> higher prio queue or just simply make the higher prio queue size 
> twice that of the lower prio;
> 
> For UDP, it shouldnt make a difference but for TCP you should notice some
> changes.
> 
> cheers,
> jamal
> 
>