Distributed services across the network from edge to core

The current internet architecture is evolving from a simple carrier of bits to a platform able to provide multiple complex services running across the entire Network Service Provider (NSP) infrastructure. This calls for increased flexibility in resource management and allocation to provide dedicated, on-demand network services, leveraging a distributed infrastructure consisting of heterogeneous devices. More specifically, NSPs rely on a plethora of low-cost Customer Premise Equipment (CPE), as well as more powerful appliances at the edge of the network and in dedicated data-centers. Currently a great research effort is spent to provide this flexibility through Fog computing, Network Functions Virtualization (NFV), and data plane programmability. Fog computing or Edge computing extends the compute and storage capabilities to the edge of the network, closer to the rapidly growing number of connected devices and applications that consume cloud services and generate massive amounts of data. A complementary technology is NFV, a network architecture concept targeting the execution of software Network Functions (NFs) in isolated Virtual Machines (VMs), potentially sharing a pool of general-purpose hosts, rather than running on dedicated hardware (i.e., appliances). Such a solution enables virtual network appliances (i.e., VMs executing network functions) to be provisioned, allocated a different amount of resources, and possibly moved across data centers in little time, which is key in ensuring that the network can keep up with the flexibility in the provisioning and deployment of virtual hosts in today’s virtualized data centers. Moreover, recent advances in networking hardware have introduced new programmable network devices that can efficiently execute complex operations at line rate. As a result, NFs can be (partially or entirely) folded into the network, speeding up the execution of distributed services. The work described in this Ph.D. thesis aims at showing how various network services can be deployed throughout the NSP infrastructure, accommodating to the different hardware capabilities of various appliances, by applying and extending the above-mentioned solutions. First, we consider a data center environment and the deployment of (virtualized) NFs. In this scenario, we introduce a novel methodology for the modelization of different NFs aimed at estimating their performance on different execution platforms. Moreover, we propose to extend the traditional NFV deployment outside of the data center to leverage the entire NSP infrastructure. This can be achieved by integrating native NFs, commonly available in low-cost CPEs, with an existing NFV framework. This facilitates the provision of services that require NFs close to the end user (e.g., IPsec terminator). On the other hand, resource-hungry virtualized NFs are run in the NSP data center, where they can take advantage of the superior computing and storage capabilities. As an application, we also present a novel technique to deploy a distributed service, specifically a web filter, to leverage both the low latency of a CPE and the computational power of a data center. We then show that also the core network, today dedicated solely to packet routing, can be exploited to provide useful services. In particular, we propose a novel method to provide distributed network services in core network devices by means of task distribution and a seamless coordination among the peers involved. The aim is to transform existing network nodes (e.g., routers, switches, access points) into a highly distributed data acquisition and processing platform, which will significantly reduce the storage requirements at the Network Operations Center and the packet duplication overhead. Finally, we propose to use new programmable network devices in data center networks to provide much needed services to distributed applications. By offloading part of the computation directly to the networking hardware, we show that it is possible to reduce both the network traffic and the overall job completion time.