[e2e] Re: Traffic Management in real networks

Matthew Roughan roughan at research.att.com
Tue Dec 30 07:10:04 PST 2003

Actually, we found that on real problems, with real IP traffic, and a 
real network, the marginal gain using a multicommodity flow solution to 
get the optimal routing was at best ~12%. That's comparing the optimal 
solution with real traffic, rather than estimated which we use in the 
OSPF technique. If one uses estimated traffic in both, with the errors 
introduced thereby, the OSPF approach was somewhat better, as it was 
more robust to those errors.

Now, that is not to say that one cannot design a robust multi-commodity 
flow optimization approach that will do as well as OSPF weight 
optimization, but we have shown fairly conclusively that there is only a 
small possible gain one can make in doing so.

Have a look at
     "Traffic Engineering with Estimated Traffic Matrices",
         ACM SIGCOMM Internet Measurement Conference 2003.

and the references therein. There are some other practical advantages of 
the approach, as listed in the paper (I won't go into details here).


P.S. note the emphasis on real networks and traffic. One can always 
concoct pathological scenarios where OSPF approaches will do badly. The 
question is how well they do in real, practical scenarios.

Marcel Waldvogel wrote:
> By leaving the limiting realms of traffic engineering by tuning OSPF 
> weights, the gain in flexibility and thus utilization can be 
> significant. Two examples include:
> Load-Sensitive Routing of Long-Lived IP Flows (Shaikh, Rexford, Shin, 
> SIGCOMM '99, http://www.acm.org/sigcomm/sigcomm99/papers/session6-2.html)
> Profile-Based Routing and Traffic Engineering (Suri, Waldvogel, Bauer, 
> Warkhede, Computer Communications 24(4), 
> http://marcel.wanda.ch/Publications/suri03profilebased)
> -Marcel
> Matthew Roughan schrieb:
>> As a follow-up, AT&T's work on estimating traffic matrixes from link 
>> loads can be found at
>>     http://www.research.att.com/projects/tomo-gravity/
>> Note that there is a paper on traffic engineering using these estimates.
>>     "Traffic Engineering with Estimated Traffic Matrices",
>>        ACM SIGCOMM Internet Measurement Conference 2003.
>> I know Sprint have some similar estimation work -- see the URL below.
>> Matt
>> Jennifer Rexford wrote:
>>>> Is there any body could do me a favor to give any information
>>>> on this topic
>>> See papers at recent SIGCOMM, SIGMETRICS, and Internet Measurement
>>> Conferences for papers on the techniques that Sprint and AT&T use
>>> for traffic matrix estimation to drive their traffic engineering tools.
>>> Also, the two papers
>>>   "Traffic engineering with traditional IP routing protocols"   (from 
>>> the October 2002 issue of IEEE Communication Magazine)
>>>     http://www.research.att.com/~jrex/papers/ieeecomm02.pdf
>>>   "Guidelines for interdomain traffic engineering"
>>>   (from the October 2003 issue of ACM Computer Communications Review)
>>>     http://www.research.att.com/~jrex/papers/ccr03.pdf
>>> give a view of how ISPs running OSPF/IS-IS and BGP tune the 
>>> configuration of these routing protocols to control the flow of 
>>> traffic in response to congestion and failure, and in preparation for 
>>> planned maintenance.  See also the tools for intradomain traffic 
>>> engineering described in
>>>   http://www.cariden.com (Cariden MATE framework)
>>>   http://www.opnet.com (OpNet SP Guru)
>>>   http://www.research.att.com/~jrex/papers/ieeenet00.pdf (AT&T Netscope)
>>> See also the papers on traffic engineering at
>>>   http://ipmon.sprintlabs.com/
>>> -- Jen

More information about the end2end-interest mailing list