Wireless Device-to-Device Caching Networks: Basic Principles and System Performance

Mingyue Ji; Giuseppe Caire; Andreas F. Molisch

Source-linked AI summary

Wireless Device-to-Device Caching Networks: Basic Principles and System Performance

Mingyue Ji, Giuseppe Caire, Andreas F. Molisch

arXiv:1305.5216v2May 22, 2013cs.ITcs.MMcs.NI

TL;DR

The paper addresses spectrally efficient video-on-demand streaming by studying device caching with D2D communications. It compares this approach with base-station-only schemes and finds competitive performance, including under realistic propagation conditions.

Problem
Video-on-demand streaming requires spectrally efficient delivery because conventional unicasting throughput decreases linearly with the number of users, while requests exhibit asynchronous content reuse.
Method
The paper evaluates D2D caching networks across throughput-outage tradeoffs and realistic propagation conditions, including microwave and millimeter-wave links.
Results
The D2D caching network shows superiority over competing schemes even under realistic propagation conditions that include factors limiting D2D communications.
Takeaways & Limitations
D2D caching delivers competitive performance while requiring simple decentralized caching without sophisticated network coordination.
Takeaways & Limitations
The evaluation includes propagation and link constraints such as NLOS propagation and limited link conditions that can limit D2D communications.

Abstract

from arXiv · show

As wireless video transmission is the fastest-growing form of data traffic, methods for spectrally efficient video on-demand wireless streaming are essential to service providers and users alike. A key property of video on-demand is the asynchronous content reuse, such that a few dominant videos account for a large part of the traffic, but are viewed by users at different times. Caching of content on devices in conjunction with D2D communications allows to exploit this property, and provide a network throughput that is significantly in excess of both the conventional approach of unicasting from the base station and the traditional D2D networks for regular data traffic. This paper presents in a semi-tutorial concise form some recent results on the throughput scaling laws of wireless networks with caching and asynchronous content reuse, contrasting the D2D approach with a competing approach based on combinatorial cache design and network coded transmission from the base station (BS) only, referred to as coded multicasting. Interestingly, the spatial reuse gain of the former and the coded multicasting gain of the latter yield, somehow surprisingly, the same near-optimal throughput behavior in the relevant regime where the number of video files in the library is smaller than the number of streaming users. Based on our recent theoretical results, we propose a holistic D2D system design that incorporates traditional microwave (2 GHz) as well as millimeter-wave D2D links; the direct connections to the base station can be used to provide those rare video requests that cannot be found in local caches. We provide extensive simulations under a variety of system settings, and compare our scheme with other existing schemes by the BS. We show that, despite the similar behavior of the scaling laws, the proposed D2D approach offers very significant throughput gains with respect to the BS-only schemes.

I. INTRODUCTION

The introduction identifies asynchronous content reuse as a central challenge for on-demand video streaming and motivates caching with device-to-device communication. It presents the paper’s tutorial, scaling-law, and realistic-system analyses, including a composite microwave/mm-wave D2D design that outperforms competing schemes.

Motivation: On-demand video traffic is projected to grow by almost two orders of magnitude in the next five years, while conventional remedies offer limited gains or high implementation costs.The cited limitations include constrained practical throughput improvements, expensive deployment, and prohibitive high-speed backhaul for dense small-cell networks.
Motivation: Asynchronous content reuse prevents naive wireless multicasting from exploiting repeated popular-file requests, because users rarely request the same file within a few seconds.Conventional unicast also faces a per-user throughput that decreases linearly with the number of users.
Caching and D2D: Caching exploits content reuse despite request asynchronism and can provide significant order gains in throughput by storing video files in users’ or helper nodes’ local caches.The paper centers on user-device caching combined with short-range D2D communications, forming a common virtual cache across devices.
Caching and D2D: The common virtual cache capacity of coded multicasting and D2D caching grows linearly with the number of users, enabling gains unavailable when content is stored only in network infrastructure.Aggregate cache capacity therefore increases with network size, even when each individual device cache is limited.
Paper scope and contributions: The paper compares throughput-versus-outage tradeoffs and develops a realistic single-cell D2D system combining robust microwave links with high-capacity mm-wave links.It reports that realistic environmental conditions strongly affect throughput and outage probability, and that D2D caching largely outperforms competing schemes on both measures.

II. LITERATURE REVIEW · A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching

The literature review surveys wireless caching networks, emphasizing D2D caching while also covering BS-only caching and uncached D2D communication. It then explains how local caching and short-range D2D links can overcome conventional ad hoc throughput scaling limits.

II. LITERATURE REVIEW: The review emphasizes caching combined with D2D communications, while also considering BS-only transmission with caching and pure D2D communication.These are the three communication settings identified for comparison.
A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching: Conventional ad hoc networks with random source-destination pairs achieve per-user throughput scaling of Θ(1/√n) under the protocol model.Here, n denotes the number of network nodes or users.
A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching: The same Θ(1/√n) per-user throughput bottleneck applies to practical relaying schemes under the protocol model.This conclusion follows despite more varied results under realistic physical models with propagation pathloss and interference.
A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching: This conventional result assumes network traffic of Θ(n), corresponding to a constant requested throughput per user, and does not capture video-on-demand content reuse.Treating every streaming session as independent data causes per-user throughput to vanish as total demand grows.
A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching: Transport capacity, defined as the sum of link throughput multiplied by source-destination distance, scales as Θ(√n) in dense ad hoc networks.The stated setting has fixed area O(1), node density Θ(n), and decode-and-forward relaying under the protocol or physical model.
A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching: For random source-destination pairs at distance O(1), throughput per link again scales as Θ(1/√n), whereas one-hop delivery can achieve constant per-user throughput.One-hop delivery requires reducing source-destination distance to the minimum inter-node distance Θ(1/√n).
A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching: Achieving constant throughput requires each user to find its requested file within its neighborhood with high probability.Caching files across the network enables requests to be satisfied through short-range links.
A. Conventional Scaling Laws Results of Ad Hoc Networks and D2D communications with Caching: Short-range caching links improve throughput because many links can share the same spectrum, increasing spatial reuse linearly with the number of users.This observation motivates practical one-hop D2D transmission with video files cached at users.

B. Network Model and Problem Definitions

The section formalizes an uncoded D2D caching network with clustered single-hop communication, random caching, interference-aware scheduling, and ergodic per-user throughput. It defines outage and the achievable throughput–outage tradeoff used to evaluate cache placement, admission control, and transmission policies.

Network model: The model places n users on a unit-square regular grid, with requests drawn independently from a common Zipf distribution over m files.Simulations may instead use uniformly random nodes in a 600m×600m square.
Network model: Communication is single-hop under a protocol model: links require distance at most r and no competing transmitter within (1+∆)r of the destination.The common range r is a design parameter and transmissions occur at rate C_r bit/s/Hz, non-increasing in r.
Caching and streaming: Each user independently caches M files according to placement probabilities P_c(f), while video is delivered in chunks and evaluated by long-term average throughput.With sufficient buffering, average information throughput matches the video source coding rate and determines playback quality.
Cluster operation: Users search only within their own cluster, with at most one transmission per cluster per resource and equal-probability scheduling among admitted potential links.Clusters contain g_c(m) nodes, and a reuse scheme with parameter K prevents inter-cluster interference.
Performance definitions: An outage occurs when a requested file is unavailable in the user’s cluster or admission control rejects the request, and p_o is the average outage fraction.All admitted users receive the same minimum average throughput T_min under the scheduling policy.
Performance definitions: An outage-throughput pair (p,t) is achievable when p_o ≤ p and T_min ≥ t; T*(p) is the supremum achievable throughput and is non-decreasing in p.The optimization ranges over cache placement, admission control, and transmission scheduling policies.

C. Key results for D2D networks with caching

The section characterizes optimal random caching and the outage-throughput scaling of clustered one-hop D2D networks in the small-library regime. The dominant scaling term is accurate in finite systems, and the resulting tradeoff is tight to the order of its dominant terms.

Optimal random caching: Theorem 1 characterizes the optimal random caching distribution under the stated model and clustering transmission scheme.It maximizes the probability that a user finds its requested file inside its own cluster.
Small-library scaling: Theorem 2 gives the achievable outage-throughput tradeoff for one-hop D2D networks with random caching and clustering transmission in the small-library regime.The regime concerns file-library scaling with the number of users.
Practical regimes: The dominant term in the tradeoff accurately captures finite-dimensional system performance, while the first two regimes are most relevant for small outage probability.The third and fourth regimes correspond to large outage probabilities that asymptotically approach 1 as m →∞.
Practical regimes: The first regime uses a large cluster size gc(m), whereas the second has m∗< m when gc(m) < γrm/M.In the first regime, m∗= m and all files are stored in the common virtual cache with positive probability.
Optimality: Theorem 2’s throughput-outage scaling laws are tight because an upper bound for any one-hop protocol-model scheme has dominant terms of the same order.The upper bound differs only slightly in its terms from the achievable result.

D. Coded Multicasting From the Base Station

Coded multicasting uses deterministic sub-packetized cache placement and base-station linear combinations to satisfy arbitrary demands with zero outage. The scheme is information-theoretically optimal within a bounded multiplicative gap of 12 under the stated protocol model, but deterministic placement requires rearranging caches when the library or user count changes.

Cache placement and delivery: Each cache stores a fraction M/m of packets from every library file, enabling the base station to multicast coded linear combinations that serve arbitrary requests.The coded messages combine packets from requested files, allowing users to recover their missing packets from cached content.
Performance guarantees: Zero outage probability is guaranteed for arbitrary demands under the ideal base-station-only protocol model.The model assumes all nodes receive the same rate with zero packet error probability.
Optimality: Within a bounded multiplicative gap not larger than 12, coded multicasting is information-theoretically optimal under arbitrary demands, zero outage, and the protocol model.A cut-set argument lower-bounds the required transmissions by 1/12 of the scheme’s benchmark quantity.
Limitations and adaptation: Deterministic cache placement achieves the scheme’s throughput but requires all caches to be rearranged when the library size or user count changes.Randomized placement was proposed to address changing user populations; its clique-covering delivery can be solved greedily within a constant factor of optimality.

E. Harmonic and Conventional Broadcasting · F. Summary: Comparison between Different Schemes

Harmonic broadcasting partitions and periodically transmits video blocks across parallel downlink channels, guaranteeing playback startup within τ chunks while achieving an average per-user throughput of R (1 −po). Across the compared schemes, conventional unicasting and local-only caching scale as Θ(1/n), whereas uncoded D2D caching and coded multicasting achieve substantially stronger scaling in the stated regime, with practical performance depending on propagation and user-channel conditions.

E. Harmonic and Conventional Broadcasting: Harmonic broadcasting splits each video into successive block sets and repeats each set periodically on a parallel downlink channel.The i-th set contains i blocks of length τ/i and is transmitted at rate R/i.
E. Harmonic and Conventional Broadcasting: Users receive harmonic-broadcast channels in parallel and can start playback after at most τ chunks.The scheme constrains the maximum waiting delay to τ chunks.
E. Harmonic and Conventional Broadcasting: R (1 −po) is the harmonic-broadcasting average throughput per user, with outage probability po = Pm.Each file requires a downlink rate of R log(L/τ).
E. Harmonic and Conventional Broadcasting: Conventional cellular and Wi-Fi streaming handles requests exclusively at the application layer and treats them as independent data.This approach uses admission control to serve a fraction 1−po of users and deny service to outage users.
F. Summary: Comparison between Different Schemes: Uncoded D2D caching and coded multicasting have equivalent throughput scaling laws and unbounded gains over conventional unicasting and harmonic broadcasting when Mn ≫m.The comparison takes M constant and m, n, L →∞, with total network storage larger than the library size.
F. Summary: Comparison between Different Schemes: Θ(1/n) is the per-user throughput scaling for conventional unicasting without D2D or caching.The comparison considers a dense single-cell network with one base station and n user nodes.
F. Summary: Comparison between Different Schemes: Θ(1/n) is also the fundamental scaling behavior of conventional caching without D2D when users cache their M most popular files.Requested files outside the local cache must be downloaded from the base station as in the no-caching case.
F. Summary: Comparison between Different Schemes: Practical throughput and outage depend on D2D-link availability, propagation and frequency band, while coded multicasting faces a worst-case user bottleneck.The bottleneck arises because a common coded message is sent to all users at one common rate despite differing path losses and shadowing.

III. SYSTEM DESIGN · A. Holistic Multi-Frequency D2D System Design

The system combines multi-frequency D2D communication with cellular fallback to exploit short-range mm-wave delivery while maintaining service when links or cached files are unavailable. It uses clustered devices and randomized complete-file caching, with cluster sizing determined by the target D2D outage probability.

III. SYSTEM DESIGN: The model excludes sub-1GHz LTE bands because their availability is not universal.The BS-to-MS assumption uses a standard LTE band at 2.1 GHz.
A. Holistic Multi-Frequency D2D System Design: The delivery hierarchy prioritizes mm-wave D2D, then 2.45 GHz D2D, and finally BS cellular downlink when local delivery fails.Mm-wave links are efficient for short-range file delivery but can be blocked by walls or the human body; BS service depends on admission control.
A. Holistic Multi-Frequency D2D System Design: D2D communication is restricted to users within the same cluster, while requests unavailable through local D2D may be served by the BS.The delivery algorithm combines D2D communications with multicast by the base station.
A. Holistic Multi-Frequency D2D System Design: The system uses independent randomized placement of complete files in device caches.Each node randomly caches M files according to the probability distribution specified by the caching design.
A. Holistic Multi-Frequency D2D System Design: Cluster size gc(m) is selected from the given D2D outage probability, and the design focuses on small outage where all potential in-cluster requests are served.The clustering and caching placement procedure is summarized in Algorithm 1.
A. Holistic Multi-Frequency D2D System Design: Randomized caching can select the same file multiple times in one cache, but this has negligible simulated impact and should be avoided in practice.The duplication does not affect throughput scaling laws, although practical caching algorithms should prevent it.

B. Conventional Unicasitng, Coded Multicasting and Harmonic Broadcasting Approaches •

The section formulates throughput–outage tradeoffs for conventional unicasting, coded multicasting, and harmonic broadcasting. Conventional unicasting achieves T_min = Θ(1/n) for any target outage probability po in (0, 1).

Conventional Unicasting Approach: Conventional unicasting caches an M/m fraction of each file and imposes equal outage probability across users.Resource allocation maximizes the minimum user rate among users not in outage.
Conventional Unicasting Approach: T_min = Θ(1/n) for any target po in (0, 1) under conventional unicasting.Varying po yields the achievable throughput–outage tradeoff.
Coded Multicasting Approach: Coded multicasting requires a common downlink rate decodable by all served users, with the largest-distance users imposing the most stringent outage condition.The throughput–outage tradeoff is obtained by varying the common downlink control parameter Cr0.
Harmonic Broadcasting Approach: Harmonic broadcasting has outage from either insufficient physical-channel quality or a request absent from the BS-broadcast file set.These independent events determine the outage probability, and varying Cr0 yields its throughput–outage tradeoff.

IV. SIMULATIONS AND DISCUSSIONS … 1) LOS probability:

The simulations evaluate cellular, microwave D2D, and millimeter-wave D2D links in office and indoor-hotspot environments using propagation models that account for path loss, shadowing, LOS conditions, and body obstruction. The channel treatment distinguishes frequency-dependent effects, including severe mm-wave sensitivity to BLOS and wall penetration, from microwave body-shadowing and indoor/outdoor scenarios.

A. Deployment Environments: The study simulates office environments and indoor hotspots within a 0.36km^2 cell containing buildings, streets, and randomly distributed users.The grid model uses n = 10000 nodes, averaging 2 ∼3 nodes per 10×10-meter square.
B. Channel Models: The simulations model cellular, microwave D2D, and millimeter-wave D2D transmissions using corresponding channel models based mostly on Winner II.Only path loss and shadowing are modeled because small-scale fading is assumed removable through frequency/time diversity.
B. Channel Models: The channel models’ validity includes device heights of about 1.5m, typical of user-held devices, although they are not explicitly defined for D2D.This limitation is noted when adapting the Winner II models to device-to-device links.
1) LOS probability:: LOS probability governs path loss, delay spread, and angular spread, but carrier-frequency-dependent subtleties affect overall performance.The study therefore distinguishes nominal LOS from antenna-level body-obstructed LOS (BLOS), which is especially important at mm-wave frequencies.

4) Channel between the Base Station and Devices:

The base-station-to-device channel uses Winner II urban macro-cell models for outdoor links and urban macro outdoor-to-indoor models for indoor communication, with added rotational body shadowing. Simulations also assume a frequency reuse factor K to avoid inter-cell interference.

Channel models: Winner II C2 and C4 models represent outdoor-to-outdoor and outdoor-to-indoor base-station communications, respectively.The only channel modification is adding rotational body shadowing.
Body shadowing: Rotational body shadowing uses the AP2HH model with σLb = 2.3 dB for LOS and σLb = 2.2 dB for NLOS.The AP2HH model is applied to access-point-to-handheld-device links.
Pathloss and shadowing: For C2 NLOS propagation, the shadowing term χσ is zero-mean with standard deviation σ = 8 over 50m < d < 5000m.The NLOS pathloss also depends on base-station height hBS and carrier frequency fc[GHz].
Inter-cell interference: Simulations assume a frequency reuse factor K to avoid interference between cells.This assumption is included to model a realistic multi-cell scenario.

5) Link Capacity Computation: · C. Results and Discussions · 1) Throughput-Outage Tradeoff:

The paper computes link capacity from SINR, bandwidth, transmit and antenna gains, interference, and receiver noise, then evaluates throughput–outage tradeoffs under specified simulation settings. The 2.45 GHz D2D scheme substantially outperforms unicasting, harmonic broadcasting, and coded multicast, with channel diversity and propagation conditions shaping the results.

5) Link Capacity Computation:: Link capacity is determined from SINR and channel bandwidth, with received signal power depending on transmit power and antenna gains, and interference aggregated at the receiver.The noise model uses kBTe = −174 dBm/Hz and receiver noise figure FN = 6 dB.
C. Results and Discussions: Simulations use n = 10000 uniformly and independently distributed users, m = 300 library files, M = 20 cached files, and Zipf request parameter γr = 0.4.The harmonic-broadcasting configuration uses L = 2.7 Gbits and P = 540 blocks.
1) Throughput-Outage Tradeoff:: The 2.45 GHz D2D-only scheme achieves throughput markedly higher than conventional unicasting, harmonic broadcasting, and coded multicast, by an order of magnitude at low outages.This result indicates that practical performance differences extend beyond asymptotic scaling-law behavior.
1) Throughput-Outage Tradeoff:: D2D attains a favorable throughput–outage tradeoff without a base-station backstop because short-distance links provide higher capacity and channel diversity.D2D outages arise from both physical channel conditions and missing requested files in the corresponding cluster.
1) Throughput-Outage Tradeoff:: The qualitative behavior of the evaluated schemes holds in both indoor office and indoor hotspot environments.The D2D advantage in the hotspot is attributed to lower interferer LOS probability and higher useful-signal LOS probability, while coded multicast performs worse there because of greater building pathloss.
1) Throughput-Outage Tradeoff:: With theoretically derived cluster sizes, the hotspot throughput–outage tradeoff can be non-monotonic because smaller clusters increase LOS probability while useful signals may be NLOS and interferers LOS.Similar behavior appears for the indoor office model under different parameter settings.
1) Throughput-Outage Tradeoff:: The observed non-monotonicity is a consequence of deploying a cluster size derived under one specific model, rather than evidence that the practical optimum must be non-monotonic.For mm-wave communication, interference is modeled as zero because the angle of arrival is narrower than 10 degrees; receiver noise-figure differences are neglected because their performance impact is low.
1) Throughput-Outage Tradeoff:: Beyond its performance advantage over coded multicast, D2D uses simpler cache placement and delivery, whereas the coded-multicasting construction does not scale well with n.The coded-multicasting approach constructs cache contents and delivery combinatorially.

2) Holistic Multi-Frequency D2D System Performance: · 3) Effects of the Density of Nodes: · 4) Effects of the Storage Capacity and the Library Size:

The holistic multi-frequency D2D system gains substantial throughput from 38 GHz links, with cluster size controlling the throughput–outage tradeoff and the base station reducing outage. Performance also depends on user density, storage capacity, and library size: density creates a throughput tradeoff, storage improves throughput, and larger libraries reduce it.

2) Holistic Multi-Frequency D2D System Performance:: 38 GHz D2D communications significantly increase average throughput per user, while a 100m × 100m cluster serves about 30% of users above 2 Mbps.For this cluster size, only around 250 users have rates below 100 Kbps.
2) Holistic Multi-Frequency D2D System Performance:: Larger clusters reduce outage probability, whereas smaller clusters produce lower minimum throughput but larger maximum throughput.A larger cluster increases the probability that the desired file is found locally.
2) Holistic Multi-Frequency D2D System Performance:: For a 100m × 100m cluster, almost no users receive below 100 Kbps and about 90% can obtain HD-quality service; the base station can serve 400 ∼500 users in the indoor office model.The base station’s role in this scenario is to reduce outage probability.
3) Effects of the Density of Nodes:: The throughput–outage tradeoff does not depend on user number or density when n and m are large and Mn ≫m, but practical density affects link rates.Higher density enables smaller clusters and shorter links with higher SINR, while also increasing LOS interference that can degrade performance.
4) Effects of the Storage Capacity and the Library Size:: In the regime nM ≫m, throughput depends linearly on user storage capacity M, allowing cache memory to be traded directly for throughput.With small outage, simulations show average throughput per user increasing even faster than linearly with M because larger M permits smaller clusters at constant outage.
4) Effects of the Storage Capacity and the Library Size:: For fixed cache capacity M, throughput decreases roughly inversely proportional to library size m.This result is shown in the throughput–outage tradeoff for different library sizes.
4) Effects of the Storage Capacity and the Library Size:: With a fixed cluster size, bandwidth division has no tradeoff with average throughput, whereas prioritizing outage creates a clear bandwidth–outage tradeoff, especially for small clusters.The base station can satisfy costly links that would otherwise increase outage probability or cluster size.
4) Effects of the Storage Capacity and the Library Size:: The minimum-outage bandwidth division is Bd2d/BBS = 0.2 for the office model and Bd2d/BBS = 0.1 for the hotspot model.The hotspot model’s better link rate explains its lower optimal D2D bandwidth allocation.

V. CONCLUSIONS

The paper concludes that D2D caching and coded multicasting outperform conventional schemes in throughput–outage tradeoffs, while realistic simulations show D2D caching remains competitive and efficiently exchanges device cache memory for throughput.

Scaling-law conclusions: D2D caching and coded multicasting outperform conventional schemes in the throughput–outage tradeoff under the dense-network protocol model.The model captures interference and spatial spectrum reuse through geometric link-conflict constraints.
Realistic-system evaluation: D2D caching remains superior or highly competitive under realistic propagation conditions, including NLOS propagation, limited range, environment shadowing, and human body shadowing.The simulations use a holistic design with D2D links at 38 GHz and 2.45 (or 2.1) GHz, plus a cellular downlink at 2.1 GHz.
System-design implications: The proposed D2D system efficiently trades user-device cache memory for system throughput.The conclusion motivates this trade because device memory is a growing, cheap, untapped resource, whereas throughput is scarce and expensive.
System-design implications: D2D caching requires simple decentralized caching and no sophisticated network coding to share files over D2D links.This simplicity strengthens the case for developing and deploying D2D caching networks.