Proceedings of the 2nd ACM Symposium on Cloud Computing
| Hi-index | 0.00 |
Even though emerging data-intensive networking applications are becoming increasingly heterogeneous, several of them require huge amounts of bandwidth which can be requested on-demand (the "dial-for-bandwidth" paradigm). Optical backbone mesh networks are suitable for accommodating these applications, as they provide large bandwidth, especially when employing wavelength-division multiplexing (WDM). A particular challenge faced by the network operator is to provision efficient service for these applications (requesting heterogeneous bandwidth, with service durations ranging from seconds to months), while meeting the customers' requirements. This dissertation investigates algorithms for efficient provisioning of data-intensive applications in optical backbone mesh networks. This dissertation first explores the problem of efficient on-demand service provisioning for applications aggregating large files from multiple remote sites to a central facility (considering different bandwidth granularities), which is useful for many data-intensive scientific applications. We propose a mixed integer linear program (MILP) formulation and heuristic solutions and investigate network performance for various problem scenarios. Next, we investigate the problem of provisioning data-intensive applications that require to transfer their data before predetermined deadlines. For such applications, the amount of allocated bandwidth is not a concern for the customer as long as its service deadline is met. We propose heuristic algorithms and a MILP formulation for this problem and show that our algorithms' performance depends on the traffic distribution and on the node architecture employed. We also study the problem of integrated optical network design for sliding scheduled traffic, in which the bandwidth is not necessarily needed immediately. We propose a non-linear mathematical model and an integrated heuristic algorithm coupled with Lagrangean Relaxation, which improves over an existing two-step heuristic, and compare our approaches with solutions of an integer linear program. Many high-performance applications require distribution of data to multiple destinations, which does not necessarily need to take place immediately. We propose algorithms for provisioning of sub-wavelength multicast data-distribution requests with flexible start times, including approaches that split multicast trees into subtrees with independent times, which partition the dataset, and which consider buffering. This dissertation introduces and investigates novel algorithms and architectures, and proposes practical solutions to help meet the needs of next-generation bandwidth-intensive applications.