Network-aware resource allocation algorithms for service orchestration in heterogeneous cloud environments
30 April 2019
Stadscampus, Promotiezaal van de Grauwzusters - Lange Sint-Annastraat 7 - 2000 Antwerpen (route: UAntwerpen, Stadscampus
Organization / co-organization:
Department of Mathematics and Computer Science
Steven Latré & Juan Felipe Botero Vega
PhD defence Bart Spinnewyn - Faculty of Science, Department of Mathematics and Computer Science
The next generation of Internet services, e.g., self-driving cars, augmented reality and cloud robotics, requires ultra-low latency wireless communications, produces vast quantities of data and requires the speedy deployment of real-time collaborations between a wide variety of devices. While traditionally cloud services are hosted on infrastructure that is located within a single data center, these novel services have throughput and response time requirements that dictate that at least some computational tasks are executed near the location of the end user.
Centralized and geo-distributed cloud environments are worlds apart. Compared to centralized clouds, geo-distributed cloud environments are much more heterogeneous. These environments incorporate both infrastructure in data centers and infrastructure at the network edge with very limited capabilities and that is much more failure-prone. This spread on capability; reliability; connectivity; and proximity to the end-user, severely complicates the management. This thesis investigates the challenges related to the orchestration of network services across heterogeneous cloud environments and proposes novel management approaches that address these challenges.
First, we investigate how to effectively replicate data across storage nodes in these environments. We approach this problem as a runtime revenue problem, that considers both Service Level Agreements (SLAs) regarding durability and the cloud characteristics. This approach builds on a dynamic availability model that considers both the impact of failure distribution and recovery times on data loss. Second, we investigate how to protect stateless network services against a combination of node and link failure in these environments. We approach the problem of placing applications while guaranteeing a minimum availability for each application and minimizing the placement cost as a resource allocation problem. To deal with the scarcity of resources at the edge and the reliability spread, our availability-aware approach introduces protections only where they are needed. Third, we investigate how to orchestrate network services in a Network Functions Virtualization (NFV) environment. We propose orchestration algorithms that can improve the acceptance ratio and placement quality through coordination of the service composition and embedding. Not only do we develop the required orchestration algorithms for an existing service model that can generate Virtual Network Function (VNF)-Forwarding Graphs (FGs) with a tree topology, we also develop a novel service model with improved applicability and develop the required orchestration algorithms.