Multi-Capacity Bin Packing with Dependent Items and its Application to the Packing of Brokered Workloads in Virtualized Environments

摘要

Providing resource allocation with performance predictability guarantees is increasingly important in cloud platforms, especially for data-intensive applications, for which performance depends greatly on the available rates of data transfer between the various computing/storage hosts underlying the virtualized resources assigned to the application. With the increased prevalence of brokerage services in cloud platforms, there is a need for resource allocation solutions that provide predictability guarantees in such settings, in which neither application scheduling nor cloud provider resources can be managed/controlled by the broker. This paper addresses this problem, as we define the Multi-Capacity Bin Packing with Dependent Items (MCBP-DI) problem to model the various resource allocation models adopted in such a brokered setting. The MCBP-DI problem represents a class of multi-dimensional bin packing problems, in which the amount of resources consumed by a subset of the items depends on the relationship between these items. Focusing on offering predictability guarantees to data-intensive applications, we define a sub-problem of the MCBP-DI problem, namely the Network-Constrained Packing (NCP) problem, in which the items to be packed form a connected component, and the resources consumed by any subset of these items are equivalent to the cost of the cut of that subset from the component. Our definition of the NCP problem is presented as part of our proposed cloud brokerage framework, in which the optimal mapping of brokered resources to applications is decided with guaranteed performance predictability. We prove that NCP is NP-hard, and we define two special instances of the problem, for which exact solutions can be found efficiently. We develop a greedy heuristic to solve the general instance of the NCP problem, and we evaluate its efficiency using simulations on various application workloads, and network models.