There were early attempts at load-sensitive routing that resulted in congestive failure - I don't have
all the references, but I recall the early ARPANET attempts are in there, and I remember reading an
article on Token-Ring source route bridging, where all traffic would discover the fastest route, select it,
and consequently congest it, while slower paths remain available but under-utilized. I haven't been
involved with TE, but I suspect this remains a very difficult problem to solve.
I have heard of successful multipath techniques that were much more localized - bundling of equal-cost
links between two nodes. The loading can be deterministic, since only one node is sending in each
direction, so the technique required is to evenly distribute the traffic while maintaining packet order (out-of-order
packets in a flow is not desireable, and many end-nodes won't tolerate it, even with TCP). I believe this
has been successfully done using a number of techniques, all requiring that a given "flow", once it is
assigned to a path, always follow that path, so the hard part here is getting the distribution balanced without
a lot of re-distributing of flows to paths (I'm sure there are other issues as well).
Cyndi
-----Original Message-----
From: Jing Shen [mailto:jshen@cad.zju.edu.cn]
Sent: Monday, April 29, 2002 7:25 AM
To: irtf-rr@puck.nether.net
Cc: routing-discussion@ietf.org
Subject: question about multipath routing applicability and internet traffic pattern
Hi,
I want to make clear some problems while doing with multipath routing.
At current time routing is done with unipath ( including ECMP on it), while in some ISP network
network optimization is done according to the traffic pattern derived from observation.
it is found such methodology works fine in current network environment ( Am i right with current situation? )
But TE asks us to make full use of network resource and improve the network's adapting ability
to bursty traffic. Although overprovision is adopted to deal with such requirements, it is not the
least cost solution ( I think). But as we simulate some of the multipath routing method for
intradomain routing, under uniform traffic pattern multipath always shows bad congestion
delivery capacity. :-(
My questions raise with such experience is :
1. How is the network traffic distributed when considering edge-to-edge connections ? What about aggregation?
what I want to make clear is : how non_uniformility is ? what is the flow level properties ?
2. Is it reasonable to simulate a network with only a subset of node exchange traffic? How could
load sensitive routing affect the traffic distibution over the network? is there any paper on it?
I found i lost in the jungle as our experiments shows different result showed by some papers I read.
Would you please give some help?
great thanks to each word you give.
Regards
Jing Shen
This archive was generated by hypermail 2b29 : Mon Aug 04 2003 - 04:10:04 EDT