From: Howard C. Berkowitz (hcb@gettcomm.com)
Date: Wed Oct 16 2002 - 16:06:20 GMT-3
At 2:23 PM -0400 10/16/02, Peter van Oene wrote:
>In many networks, particularly ISP networks, the topologies are not
>such that there is lot of metric parallelism in paths and equal cost
>load balancing alone does not effectively utilize a given topology.
>Furthermore, it may not be that you want to purely load balance
>traffic, but you might want to selectively map certain traffic to
>certain links and other traffic to other links. Doing this with
>metric manipulation is not possible, and policy routing needs to be
>done per router along a given path and really has some provisioning
>drawbacks, not too mention may seriously degrade forwarding
>performance on some platforms.
Perhaps this is a rewording of what Peter is saying, but experience
and research have shown that link-by-link load balancing winds up
producing route table oscillation, and, depending on the balancing
algorithm, may produce pinhole congestion or significant
out-of-sequence errors.
Some of the best early work was done in Curtis Villamizor's Optimal
Multipath Routing work. Since he's moved from the defunct ANS, you'd
have to hunt for them. There's also good work in Vern Paxson's
dissertation, "End-to-End Packet Dynamics in the Internet."
The bottom line is that effective load balancing really has to
consider the end-to-end path, not any link on it. Otherwise, using
an apparently uncongested link may take you to a congested,
low-bandwidth next hop.
>
>These types of issues, be they QoS centric or simply a desire to
>intelligently engineer traffic flows over a given topology, are some
>of the reasons why MPLS finds use in many networks. Most people
>think of MPLS these days as a VPN enabler, but it certainly has
>other uses.
>
>At 02:16 PM 10/16/2002 -0400, Larson, Chris wrote:
>
>>Ahh well. Now your talking about real world stuff. In that case I
>>would not want to do policy routing either.
>>
>>Just out of curiosity, in the real world why would you want to take
>>a certain path on links that could LB anyway?
>>-----Original Message----- From: Peter van Oene
>>[SMTP:pvo@usermail.com] Sent: Wednesday, October 16, 2002 2:09
>>PM To: Larson, Chris; 'Volkov, Dmitry (Toronto - BCE)'; Dave
>>Cooper Cc: CCIE Subject: RE: OSPF route manipulation
>>
>>I suppose you could butcher things that way. My point of view it
>>skewed toward reality :-) Policy routing per destination on your
>>GSR is not really a good thing.
>>
>>At 01:37 PM 10/16/2002 -0400, Larson, Chris wrote: >Without
>>getting into it to much, can't you just use a route-map on the
>>>interfaces of R4 to define the wieght of the routes? Or policy
>>routing that >sets the next hop based on destination? > > >
>>-----Original Message----- > > From: Volkov, Dmitry (Toronto -
>>BCE) [SMTP:dmitry_volkov@ca.ml.com] > > Sent: Wednesday, October
>>16, 2002 1:05 PM > > To: Dave Cooper > > Cc: 'Peter van
>>Oene'; CCIE > > Subject: RE: OSPF route manipulation > >
>>> > I guess You can use distance : > >
>><http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121cgcr/ip>http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121cgcr/ip
>>> > _r > > /iprprt2/1rdindep.htm#xtocid2 > > > > > > Dmitry
>>> > > > > -----Original Message----- > > > From: Peter van Oene
>>[<mailto:pvo@usermail.com>mailto:pvo@usermail.com] > > > Sent:
>>Wednesday, October 16, 2002 12:32 PM > > > To: Dave Cooper; CCIE
>>> > > Subject: Re: OSPF route manipulation > > > > > > > > >
>>Can't do it. This type of granular traffic engineering is one of
>>the > > > things MPLS-TE is useful for. Metric based TE is
>>topology > > > centric and > > > fixed for all destinations over
>>a given topology. The > > > topology itself > > > cannot be
>>modified on a per destination basis. > > > > > > At 06:14 AM
>>10/16/2002 -0700, Dave Cooper wrote: > > > >Hi > > > >
>>> > > >Four routers in a diamond shape > > > > > > > >
>>|----r2----| > > > >r1 r4 > > > > |----r3----| > > > >
>>> > > >All are in area 1. All interfaces serial p-2-p > > > >
>>> > > >R1 is advertising 2 routes x.x.0.0 & y.y.0.0 > > > >R4 has
>>2 equal paths for each (via r2 & via r3) > > > > > > > >How can I
>>force r4 to "prefer" r2 for x.x.0.0 and r3 > > > >for y.y.0.0
>>> > > >In case prefered path is not there, use the other
>>> > > >path. > > > > > > > >I tried using
>>cost/bandwidth/distribute-list. But they > > > >can't distinguish
>>x from y and change metric > > > >accordingly. > > > > > > > >Is
>>this possible ? > > > > > > > >Thanks > > > >Dave > > > >
>>> > > >__________________________________________________
>>> > > >Do you Yahoo!? > > > >Faith Hill - Exclusive Performances,
>>Videos & More > > > ><http://faith.yahoo.com>http://faith.yahoo.com
This archive was generated by hypermail 2.1.4 : Tue Nov 05 2002 - 08:35:48 GMT-3