Resource Pooling A system exhibits complete resource pooling

Resource Pooling A system exhibits complete resource pooling if it behaves as if there was a single pooled resource. I propose ‘extent of resource pooling’ as a metric for evaluating multipath routing algorithms. Damon Wischik, UCL

Resource pooling is not just a fancy name for load balancing. • Load balancing means ‘the load on all resources is roughly equal’ • Complete resource pooling means ‘the system behaves as if there was one resource’

I am interested in the behaviour: ‘allocation of flow rates to long-lived flows’ • Kelly’s Multipath TCP allocates transmission rates so as to solve an optimization problem • I want to know: is this equivalent to solving a single-resource optimization problem? • If so, there is complete resource pooling.

What is the behaviour when there is complete resource pooling? • If there is complete resource pooling, – The network’s total capacity is shared fairly between flows – Adjusting the capacity of a single resource will rebalance traffic over all the other resources – A localized surge in traffic will distribute itself over the pool

Resource pooling does not imply balanced load nor equal flow rates. This is a good thing. • Here, three resources behave like a single pool, but one is more congested • Flow y 1 takes twice as much from the pool as y 2

Sometimes the pool is made up of only some of the resources — the bottleneck cut [See Laws, Resource pooling in queueing networks with dynamic routing, 1992, Example 3. 2. He finds generalized cut constraints of this form for a version of this problem with fixed capacity constraints and demands. ]

Typically the network will be split into several pools. This is called partial resource pooling. • Partial resource pooling is a sign that there aren’t enough flows with the right path diversity • Note that the resource pools depend on routing & capacity & traffic. With different traffic loads, there will be different pools.

To improve resource pooling, we need better path diversity. • To achieve better resource pooling, we need to give better path choices • This should shift traffic onto the less congested pool • If enough traffic can shift, the pools will merge

How to find the resource pools 1. Create a multipath routing 2. Measure the total rate ys for each flow s and the drop probabilities pj at each link j – I propose to find these by running the fluid model of m. TCP 3. Find subsets of resources where y and p satisfy a certain relationship; these are the resource pools

What help do I need to get started? What do I want to achieve? • • I need a sensible topology to experiment on I need a sensible multipath routing algorithm I need a sensible traffic demand matrix I would like a plausible picture of p 2 p traffic The resource pool algorithm works just as well with m. TCP as with a suitable congestionsensitive p 2 p protocol • I want a simulator which will identify the pools • We can then see where the multipath routing algorithm has not done very well, and see which extra paths it should give • Maybe we’ll find that p 2 p traffic is sufficient to achieve pretty good resource pooling! (with appropriate congestion control and peer choice) Probably, we only need the large flows to have path diversity, since they contribute most of the bytes on the Internet.