Azer Bestavros - Professor in the Computer Science Department at Boston University Digital Security
Date: - Location: Eurecom
In emerging settings where customers are empowered to make autonomous resource acquisition decisions, and in which infrastructure providers and customers are interested in selfishly maximizing their own utilities, resource management must be viewed through a game-theoretic (as opposed to a global optimization) perspective. In this talk, I will present such a perspective for customer workloads that are amenable to packing either over space (in a cloud computing setting) or over time (in a last-mile networking setting). In the first packing-over-space setting, motivated by the need for rational resource acquisition strategies in IaaS cloud offerings, customer workloads can be packed into fixed-capacity instances that are acquired and paid for on the fly. For this setting, I will present ?Colocation Games? (CG) ? an economically-sound framework upon which emerging cloud architectures could be implemented. CGs enable the modeling and analysis of the dynamics that result when rational, selfish parties interact in an attempt to minimize the individual costs they incur to secure the shared cloud resources necessary to support their application QoS or SLA requirements. In addition to various game-theoretic results, I will overview implementation considerations as well as results from experimental evaluation of the ?Cloud Commons? prototype. In the second packing-over-time setting, motivated by the need for rational last-mile bandwidth consumption strategies, customer workloads must be packed over time into a pre-paid, fixed-capacity shared resource (link). For this setting, I will present ?Trade and Cap? (TC) ? an economics-inspired mechanism that incentivizes users to voluntarily coordinate their consumption of a shared resource so as to converge to what they perceive to be an equitable allocation, while ensuring efficient resource utilization. Under TC, rather than acting as an arbiter, providers act as enforcers of what a community of rational users decides is a fair allocation. In addition to presenting the analytical underpinnings of TC and results from trace-driven simulations, I will briefly discuss implementation considerations for last-mile bandwidth arbitration. This work was pursued at Boston University in collaboration with Jorge Londono (now at Universidad Pontificia Bolivariana in Colombia), and with contributions from Vatche Ishakian, Shanghua Teng (now at USC), and Nikos Laoutaris (at Telefonica Research).