Good work, Gaz. Still, I wouldn't necessarily go too much farther in, for example, playing with the K values.
Load balancing, on a per-hop basis, is a very limited solution. Paradoxically, by considering actual load (the K value), you may make it worse. Let's first look at the problem of load balancing in the simple case: two routers with two parallel links between them. Per-packet load balancing clearly makes the optimal use of bandwidth, but it also creates the greatest potential for packet misordering, which may eventually raise processing load significantly on the destination _hosts_. As implemented by Cisco, it also makes the greatest demand on router processing. It does have the additional benefit of fastest convergence after failures. Per-destination load balancing can be quite effective if you have a large number of destinations and a large number of paths. Otherwise, it can cause "pinhole congestion," when most of your traffic goes to a single path. Source-destination hash is probably the best compromise, but is not always available. Now, consider what happens if you play with the K values. Traffic will prefer the less loaded path...until your traffic pattern changes. Then the less loaded path may become more loaded, until some traffic moves to the other path. In other words, you incur a lot of oscillation in route selection, with complex interactions with the load balancing method in use. In general, most routing protocol designers consider classical routing protocol load balancing to be an evolutionary dead end. There are several alternatives, such as experimental routing protocols that consider the load on every link in the end-to-end path. The consensus, however, tends to be to use the routing protocol to disseminate reachability information, and possibly to disseminate such things as reserved bandwidth per interface. For the timing-sensitive traffic, traffic-engineered MPLS paths will be set up with RSVP-TE. Internet- and multiprovider traffic engineering is beyond the scope of this introduction. >"Gaz" wrote, >The information turned up quicker than I thought it might. >The traffic is balanced based on the ratio of metrics. Seems a bit crude. If >the ratio is not an integer value then the value is rounded down to an >integer value. > >This suggests to me that if the metric of the lesser link is not at least >twice that of the better link, that the links will be load balanced as >though they are equal. > >All of the above is based on the fact that K values are left as default, so >load is not taken in to consideration with EIGRP. I've never messed with the >K values yet, but I think I may have to in the near future. >I think there will have to be a fair bit of testing involved once the load >is taken into configuration, but I think if done correctly, this should >offer better intelligent load balancing. > >URL to have a look at: > >http://www.cisco.com/warp/public/103/19.html > > >Gaz > > > >http://www.cisco.com/warp/public/103/19.html >""Gaz"" wrote in message >[EMAIL PROTECTED]">news:[EMAIL PROTECTED]... >> Hi Gregg/All, >> >> Would the variance have any effect on the actual load balancing in that >way? >> I was under the impression that the variance would allow the lesser route >to >> be added to the routing table, but after that the variance would have no >> effect on the number of packets sent over each link. >> Not sure, but I have an interest in finding out, so I'm off to have a >look. >> >> My thoughts at the moment are that packets may be distributed across >unequal > > routes due to the metrics of those routes, but that the variance would not >> matter whether it were 2 or 22, only that it allows the routes to be used. >> >> I'll get back to you when/if I find out. >> >> >> Gaz >> >> >> >> ""Gregg Malcolm"" wrote in message >> [EMAIL PROTECTED]">news:[EMAIL PROTECTED]... >> > I have a question about variance that's been bugging me. I know that >> metric >> > based routing proto's (IGRP, EIGRP and OSPF) will not load balance >across >> > unequal cost links by default. We must use that variance cmd. The >> variance >> > has a multiplier. 1 is equal cost. I assume that variance is done per >> > packet (as opposed to session). Is this true? If variance is set to 2 >> does >> > it mean 2 packets would be sent out high bandwidth link and 1 packet out >> the > > > low bandwidth link? Message Posted at: http://www.groupstudy.com/form/read.php?f=7&i=27925&t=27882 -------------------------------------------------- FAQ, list archives, and subscription info: http://www.groupstudy.com/list/cisco.html Report misconduct and Nondisclosure violations to [EMAIL PROTECTED]
