[e2e] Hiccups and scheduling in mobile networks

Detlef Bosau detlef.bosau at web.de
Sun Jan 7 16:05:44 PST 2007

Joe Touch wrote:
>> Variations in delivery times can be handled via PEPs that don't spoof
>> ACKs, e.g., ones that pace the data and/or ACK paths, but don't actively
>> participate in the communication.

And my humble comment was:

> Really? I agree with you for the Remote Socket Architecture 
> (Schlager/Wolisz) because that architecture actually does not split 
> the connection but places the PEP mechanism at the application/socket 
> interface.
> Otherwise the problem is: When the bandwidth sender - splitter is, 
> e.g., the average bandwidth / rate splitter-sender but far less than 
> the maximum rate splitter / sender than a simple router perhaps would 
> hardly store any data and thus hardly equalize the rate / delivery times.
> Thierry describes delay spikes of several seconds. If we think about 
> UMTS, we can imagine a wireless link were nothing happens for up to 
> several seconds - thus even no data is clocked out from the sender - 
> and then we have about 2 Mbps throuhput for a short time - which is 
> perhaps much more than the actual Internet path can carry. In such a 
> scenario we want to have the router / splitter / PEP / whateverbox 
> buffer the data and equalize the rate variations. Can this be achieved 
> by pure pacing in the one or other direction?
> Detlef

O.k. So, I see: Splitting is unsellable :-) So the question is whether 
we really need it.

So, this weekend I spent some time adding hiccups to a quite complex 
network scenario:

Sender-----(internet)--------BS----(mobile net)-------Receiver

And the mobile net suffers from hiccups :-)

What I would like to know (and AFAIK Andreas dealt with questions like 
these, therefore I put him on the cc: list) is "how bad" this hiccups my 
become. As I said before, Thierry Klein published a paper at Globecom 
2004 on this issue.  There, he observed delay spikes from up to two seconds.

For the moment, I simply model the wireless link as a link with a 
constant high bandwidth (e.g. 10 Mbps) which reflects its _physical_ rate
and I add hiccup times to the serialization delay (i.e. txtime in NS2). 
These are drawn from a two point distribution: Either the hiccup time is 
zero or it is 1 second. The probabilities are chosen that way that a 
given average throughput is achieved.

Of course that´s extremely simplified. However: Is this reasonable as a 
first approach? I would appreciate any comment on this one.

I would like to study different pacing techniques in this scenario, 
_intendedly_ without splitting.

AFAIK, there is a variety of scheduling algorithms available for 
networks like GPRS or UMTS. So, the question is whether we have a, if 
extremely rough, "worst case model" to get a feeling for what TCP has to 
cope with. The idea of my model above is to insert constant, say 1 
second, delay spikes randomly into the flow, just in a way that I can 
estimate the average throughput on the link.

Is this completely weird? Or does it sound reasonable?



More information about the end2end-interest mailing list