XPS: CLCCA: Allocating Heterogeneous Datacenter Hardware to Strategic Agents

Duke University, Durham NC

Investigators

Abstract

With the democratization of cloud computing, increasingly diverse software tasks demand computation from cloud computing. Scalable datacenter computing is a 21st century imperative and energy efficiency is the most pervasive challenge. Increasing capability to extract value from the data deluge requires order-of-magnitude gains in efficiency. While heterogeneous hardware systems provide efficiency, they also complicate resource management. Managing new hardware architectures with economic mechanisms is a qualitatively new approach to mitigate this complexity. The investigators propose resource allocation mechanisms that encourage truthful participation in datacenters with heterogeneous hardware. Resource allocation pairs hardware with software. Users may explicitly request specific amounts and types of hardware. Or users may define their value for performance, which agents then use to bid for hardware. Both allocation mechanisms rely on the truthful representation of preferences from users. The investigators will quantify inefficiencies arising from strategic behavior in datacenters and will design allocation mechanisms that are robust when strategic agents participate in the system. Women and under-represented students will benefit from this research experience.