SAMBA: Scalable Approximate Forwarding For NDN Implicit FIB Aggregation

TL;DR

The paper addresses FIB scalability in NDN by introducing SAMBA, which combines Approximate Forwarding (AF) with Implicit Prefix Aggregation (IPA) to forward via the nearest FIB record and implicitly compress the FIB. It integrates a light-weight DFS-based fallback and a Stop-and-Wait consumer mechanism to minimize unnecessary discoveries, and adds multipath discovery to improve resilience while curbing flooding. Experimental results in ndnSIM show substantial gains: up to twenty-fold reductions in FIB size and significantly lower discovery overhead, with up to fifty percent higher throughput during link failures compared to on-demand Self Learning approaches. Overall, SAMBA provides scalable, resilient forwarding for ICN/NDN with reduced memory and processing burdens and practical applicability to ISP-like topologies; future work includes real-testbed deployment and refined multipath management.

Abstract

The Internet landscape has witnessed a significant shift toward Information Centric Networking (ICN) due to the exponential growth of data-driven applications. Similar to routing tables in TCP/IP architectures, ICN uses Forward Information Base (FIB) tables. However, FIB tables can grow exponentially due to their URL-like naming scheme, introducing major delays in the prefix lookup process. Existing explicit FIB aggregation solutions are very complex to run, and ICN on-demand routing schemes, which use a discovery mechanism to help reduce the number of FIB records and thus have shorter lookup times, rely on flooding-based mechanisms and building routes for all requests, introducing additional scalability challenges. In this paper, we propose SAMBA, an Approximate Forwarding-based Self Learning, that uses the nearest FIB trie record to the given prefix for reducing the number of discoveries thus keeping the FIB table small. By choosing the nearest prefix to a given name prefix, SAMBA uses Implicit Prefix Aggregation (IPA) which implicitly aggregates the FIB records and reduces the number of Self Learning discoveries required. Coupled with the approximate forwarding, SAMBA can achieve efficient and scalable forwarding

Figures (9)

• Figure 1: Scalability assessment of FIB lookup time; average of 100 FIB lookup and insertion times for as the FIB table size scales from 1k to 1M entries (x axis in logscale)
• Figure 2: Comparing the FIB properties using state-of-the-art solutions shi2017broadcast (left) and SAMBA (right). SAMBA inserts very few prefix entries, often longer but act as an implicit aggregate/summarization for the entire trie branch.
• Figure 3: Trie example: (I1) AF performs similar to LPM and returns $f_1$; (I2), AF runs a DFS on the sub-trie rooted at B and returns $f_1$ as it finds leaf node D; (I3) AF runs DFS on the sub-trie rooted at H which fails to find a next hop, thus AF returns a NOROUTE.
• Figure 4: Flowchart diagram of SAMBA's Stop-and-Wait mechanism. Interests are queued waiting for the creation of a new route or a timeout of a current active discovery.
• Figure 5: Comparison of SAMBA's and Self Learning's average FIB table size as the number of prefixes, $C$, increases. The improvement ratio, in the right y-axis, increases as the number of prefixes per producer increases (x axis in logscale).