I have trouble understanding the concept of bisection bandwidth in high performance computing clusters.
Wikipedia article on bisection bandwidth, and online materials say that bisection bandwidth is the minimum number of links one have to cut to partition the network in two equal parts.
My understanding from reading research papers in networking (like MapReduce) keep saying that bisection bandwidth represents the "lower bound" on the bandwidth that the network can "push through". Is it a correct intuition? If yes, how do I reconcile it with the definition from Wikipedia et al.?
From a perspective of one single node, does bisection bandwidth roughly represent the average bandwidth that the node can have to any other node in the network?
Best Answer
This is one of those terms for which multiple definitions have been created. Also, remember that Wikipedia is maintained by everyone, including you. It can often end up with a definition from an individual's perspective, especially with relatively obscure topics, such as this. You, or anyone else, is free to edit or add to Wikipedia articles.
Most people don't require bisectional bandwidth to have two equal parts; you can use bisectional bandwidth on any network which has been arbitrarily bisected, even if the parts are not equally sized. It may be that the Wikipedia author believes bisecting means creating two equal parts, but it only means creating two parts, equal or not, from a whole.
I think the author of Bisectional Bandwidth. And why L2MP and Trill/RBridges is important? does a good job of explaining it by using it to demonstrate a problem.
What is usually meant by bisectional bandwidth is the available bandwidth between two parts of a network. This can be created by STP blocking links that can't be used to move traffic except in a failure of a forwarding link. By using something other than STP, you may be able to suddenly double the bisectional bandwidth.
The term is not normally applied to individual nodes. Instead, it applies to links between the switches.