Rigorous Systems Research Group (RSRG) Seminar

While most people have an intuitive notion of "fairness" in allocating resources, it is difficult to quantify this intuition in a mathematical fairness function. Mathematical characterizations of fairness are particularly under-explored when there are multiple types of resources and when users combine the resources in different ratios. A typical example is datacenters processing computational jobs with heterogeneous requirements for CPU, memory, network, bandwidth, etc.

In this talk, I will develop a unifying framework for multi-resource fairness that generalizes single-resource fairness measures to address heterogeneity in user resource requirements and tradeoffs between fairness and efficiency. I will introduce two families of fairness functions that provide different fairness-efficiency tradeoffs, characterize the effect of user heterogeneity, and prove conditions under which these fairness functions satisfy the Pareto efficiency, sharing incentive, and envy-free properties.  I will also show how these functions can be mapped to people's intuitive fairness perceptions by examining the results of an online survey of allocation preferences.