[e2e] Two questions on simulation of AQM
jinw at comp.leeds.ac.uk
Wed Apr 11 03:59:20 PDT 2001
For marking situation, I think the result is same.
For example, as in ur scenario, a burst flow will very hopefully let all
the packets in queue being marked, which means one burst flow will cause
most of the flows back-off. The consequent is as follow:
1. Burst traffic is not easily to be marked.
2. Most flows back-off will cause performance oscillation.
3. Sources with burst behavior will gain a better throughput.
I think the tendency towards burst flow is unfair. But this is just my
opposite idea, we need simulations to prove it is right or wrong.
About ur second question, I agree that reverse channel does impact to TCP
performance, which is quite obviously. I am not quite clear about what u
mean "policy in the reverse direction", is it ACK sending policy or queue
management policy or switch-node scheduling policy.
The paper I mentioned in last email can solve the first situation. If u
want to provide a different queuing and scheduling policy between "sending
direction" and "reverse direction", I don't think it is worthy in my deep
mind. That will make switch-node to be too complex to operate.
School of Computing,
University of Leeds,
Leeds, LS2 9JT,
Tel: (44)113 2336806
On Wed, 11 Apr 2001, Zhang Miao wrote:
> Hi, Wu
> Your argument seems reasonable for dropping. If a packet is dropped earlier,
> it may provide a room for another packet. There is indeed a tradeoff between
> feeding back faster and the hazard of dropping more packets. But how about marking?
> Is there also such tradeoff?
> For question 2, I agree that the effect is slight. But this effect becomes more
> important when the performance of two AQM schemes are very similar. In the paper of
> REM(http://netlab.caltech.edu), the difference of goodput between RED and REM is
> very little. So several percent effect also becomes important. Perhaps this little
> difference has no meaning in simulation results.
> >For question 1, I agree with u that, if the Marking or Dropping action is
> >in deque(), it will obviously shorten the feedback latency, but will
> >increase the potential of "Lock-Out".
> >As I think, if u drop or marking the packet from the beginning of the
> >queue, some burst traffic will monopolize the queue space (or let all the
> >packets in queue being marked), preventing other connection from getting
> >room in the queue. Simply change the Marking or Dropping action the
> >deque() will not work, it will let the burst traffic grasp will the
> >For question 2, as far as I can remember, a single ACK loss will just
> >have a slight effect to TCP performance, so reverse direction policy will
> >not change the TCP performance a lot. There have already some research
> >work being done around this. U can find the paper "The Effects of
> >Asymmetric Links on TCP Performance and Its Solutions" in JOURNAL OF
> >BEIJING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Vol.22 No.4 1999
> >(Chinese). It solves the problem when the bottleneck is in reverse
> >Jin Wu
> >School of Computing,
> >University of Leeds,
> >Leeds, LS2 9JT,
> >Tel: (44)113 2336806
More information about the end2end-interest