Wireless Multihop Device-to-Device Caching Networks

TL;DR

This paper analyzes the capacity scaling laws of wireless D2D networks with multihop communication and caching, showing that decentralized caching strategies can achieve scalable throughput under various popularity distributions.

Contribution

It characterizes the optimal per-node capacity scaling laws for multihop D2D caching networks with heavy-tailed popularity distributions, including Zipf, and proposes caching strategies for throughput improvement.

Findings

Decentralized uniform caching achieves constant per-node capacity scaling of Θ(√(M/m)).

Multihop capacity scaling outperforms single-hop caching networks.

Selective caching of popular files enhances throughput for Zipf distributions with exponent > 1.

Abstract

We consider a wireless device-to-device (D2D) network where $n$ nodes are uniformly distributed at random over the network area. We let each node with storage capacity $M$ cache files from a library of size $m \geq M$. Each node in the network requests a file from the library independently at random, according to a popularity distribution, and is served by other nodes having the requested file in their local cache via (possibly) multihop transmissions. Under the classical "protocol model" of wireless networks, we characterize the optimal per-node capacity scaling law for a broad class of heavy-tailed popularity distributions including Zipf distributions with exponent less than one. In the parameter regimes of interest, we show that a decentralized random caching strategy with uniform probability over the library yields the optimal per-node capacity scaling of $\Theta(\sqrt{M/m})$, which is constant with $n$, thus yielding throughput scalability with the network size. Furthermore, the multihop capacity scaling can be significantly better than for the case of single-hop caching networks, for which the per-node capacity is $\Theta(M/m)$. The multihop capacity scaling law can be further improved for a Zipf distribution with exponent larger than some threshold $> 1$, by using a decentralized random caching uniformly across a subset of most popular files in the library. Namely, ignoring a subset of less popular files (i.e., effectively reducing the size of the library) can significantly improve the throughput scaling while guaranteeing that all nodes will be served with high probability as $n$ increases.