In a cloud market, the cloud provider provisions heterogeneous virtual machine (VM) instances from its resource pool, for allocation to cloud users. Auction-based allocations are efficient in assigning VMs to users who value them the most. Existing auction design often overlooks the heterogeneity of VMs, and does not consider dynamic, demand-driven VM provisioning. Moreover, the classic VCG auction leads to unsatisfactory seller revenues and vulnerability to a strategic bidding behavior known as shill bidding. This work presents a new type of core-selecting VM auctions, which are combinatorial auctions that always select bidder charges from the core of the price vector space, with guaranteed economic efficiency under truthful bidding. These auctions represent a comprehensive three-phase mechanism that instructs the cloud provider to judiciously assemble, allocate, and price VM bundles. They are proof against shills, can improve seller revenue over existing auction mechanisms, and can be tailored to maximize truthfulness.

In a secondary spectrum market, the utility of a secondary user often depends on not only whether it wins, but also which channels it wins. Combinatorial auctions are a natural fit here to allow secondary users to bid for combinations of channels. In this context, the VCG mechanism constitutes a generic auction that uniquely guarantees both truthfulness and efficiency. There also exists related auction design that relaxes efficiency due to perceived complexity issues, and focuses on truthfulness. Starting with new empirical evidences on the complexity issue, we propose to design core-selecting auctions instead, which resolve VCG's vulnerability to collusion and shill bidding, and improve seller revenue. While the VCG type of auctions are unique in guaranteeing both efficiency and truthfulness, we prove that our core-selecting auctions are unique in guaranteeing both efficiency and shill-proofness, and always outperform VCG auctions in terms of seller revenue generated. Employing linear programming and quadratic programming techniques, we design two payment rules for minimizing the incentives of bidders to deviate from truth telling.