Cloud Self-organization Based on Coalition Formation and Combinatorial Auctions

Dan C. Marinescu

Computer Science Division, EECS Department University of Central Florida, Orlando, FL 32816, USA

Email: dcm@cs.ucf.edu

Abstract

We first discuss existing and future challenges to resource management in large-scale systems such as computer clouds. Nowadays, the cloud infrastructure consists of multiple hierarchically-organized Warehouse Scale Computers including multi-core processors, GPUs, data-ow engines, and FPGAs. Managing such a complex system is extremely challenging and this justifies the interest in self-organization.

We present results shown that market-based mechanisms are more effective than hierarchical control when resource allocation is based on mostly obsolete information regarding the state of individual system components. Finally, we present a cloud reservation system for Big Data applications, based on coalition formation and combinatorial auctions. Resources such as CPU cycles, memory, and secondary storage provided by individual servers are often insufficient for such applications; this justifies the need for rack-level coalition formation. Moreover, many existing and future cloud applications exhibit a complex work ow with multiple phases each one best supported by server coalitions with different architectures and resources. This justifies the need for combinatorial auctions which support auctioning of packages consisting of multiple server coalitions.