Off-the-Grid Communication Support via Named Data Networking

Abstract

Most today's applications require Internet connectivity to be functional, because they are hard-coded to communicate with specific cloud servers. However, Internet connectivity can be unavailable in many scenarios, such as a remote location with no cellular coverage and during natural disasters. In these scenarios, local connectivity still exists, but applications are unable to communicate ``off-the-grid'', i.e., exchange messages without Internet connectivity. In this dissertation, we investigate both application and network support for off-the-grid communication in \textit{Named Data Networking} (NDN), a novel Internet architecture. NDN's core features (data naming, adaptive forwarding, data-centric security) make it inherently more capable of supporting applications with off-the-grid communication than the TCP/IP architecture. For example, in NDN, an application is able to retrieve and verify data from anywhere. Specifically, the investigation covers three different aspects. First, existing applications are unable to run over NDN directly. To migrate them into NDN, we propose to translate between existing application protocols and NDN, which requires a small amount of development efforts but gains significant NDN's architectural benefits. With concrete work on IMAP/NDN translation for local email access and XMPP/NDN translation for local group chat, we identify and discuss a number of common design issues. After application migration, the next challenge is to support off-the-grid communication within an NDN local network. To tackle this challenge, we utilize \textit{self-learning} mechanisms on NDN routers to help applications discover and retrieve data. The original self-learning design suffers from poor performance over wireless communication, and underutilized network resources with learning a single path. To tackle these two problems, we add mechanisms such as face creation and selection, multipath learning, and more aggressive forwarding processing. The improved self-learning is implemented in an NDN forwarder, and is evaluated on real devices. Evaluation results show the performance is significantly improved with the proposed mechanisms. The third problem of self-learning is that it uses a fixed name prefix of data name as the route name. This problem exists not only in self-learning but also in any NDN \textbf{adaptive forwarding} design, i.e., the forwarding plane observes data retrieval performance of past Interests and use it to adjust forwarding decisions of future Interests. To be effective, adaptive forwarding assumes what we call \textbf{Interest Routing Locality}, that Interests sharing the same prefix are likely to take the same or similar forwarding path within a short time window, thus past observation can be an indicator of future performance. Since Interests can have multiple common prefixes with different lengths, the real challenge is what prefix length should be used in adaptive forwarding. The longer the common prefix is, the better Interest Routing Locality, but the fewer future Interests it covers and the larger the forwarding table size. Existing implementations use static prefix length, which is known to have problems in dealing with partial network failures. In an local network with self-learning, this problem is more severe because the learned routes can point to cache, which only contains partial data. To solve this problem, we propose to dynamically aggregate and disaggregate name prefixes in the forwarding table, so to use the prefixes that are most appropriate under the current network situation. To reduce the overhead, we design mechanisms to minimize the use of longest prefix match in the processing of Data packets. Evaluation results show that our designs solve the prefix granularity problem with acceptable overhead. Because of the optimization on the measurement module, the overall processing overhead is significantly reduced.