Virtualizing Traffic Shapers for Practical Resource Allocation

Gautam Kumar, Srikanth Kandula, Peter Bodik, and Ishai Menache

Abstract

Many network resource allocation scenarios require traffic shapers, such as weighted fair queues or priority queues, to achieve their goals; however, current switches offer very few of such shapers. We show how to virtualize the traffic shapers in switches and network interface cards. Doing so implies that one can mimic the traffic shaping behavior of many more shapers (at least one order of magnitude more) by just using the small number of shapers that are available on commodity switches and NICs. From a prototype on Arista 7048 and simulations that replay traces from production datacenters, we show early results that indicate feasibility and improvement over simpler alternatives. We also present theory-based intuition as to why such virtualization of shapers is feasible.

Details

Publication typeInproceedings
PublisherHotCloud
> Publications > Virtualizing Traffic Shapers for Practical Resource Allocation