Projet de recherche doctoral numero :4765


Date depot: 1 janvier 1900
Titre: Joint Optimization of Content Caching and Recommendation for Mobile Edge Systems
Encadrant : Christian BONNET (Eurecom)
Directeur de thèse: Thrasyvoulos SPYROPOULOS (Eurecom)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini

Resumé: Summary Caching popular content at the edge of future mobile networks (e.g. small cells or user devices) has been widely considered in order to alleviate the backhaul bottleneck resulting from very dense, small cell networks, that are deemed necessary to keep up with the data tsunami experienced. While a number of interesting techniques have been proposed that target the wireless environment explicitly, such as femto-caching and coded caching, the majority of these approaches suffer from the rather limited storage capacity of edge caches. Even if one considers the additional “global” caching gains these methods promise, the envisioned gains in practical settings are rather pessimistic, due to the tremendous and rapidly increasing size of the Internet content catalog. In this project, we propose to depart from the assumption of hard cache misses, common in most of these works, and consider soft cache misses, where alternative content, related to the original one not found locally, can be recommended possibly leading to complete or partial user satisfaction, without the need to retrieve the original content over the expensive backhaul and core network. This thesis proposes a first-of-its-kind joint consideration of caching and recommendation systems, towards facilitating mobile edge caching performance. Such a convergence is quite timely due to popular content providers envisioning becoming virtual operators, facilitated by RAN Sharing and Mobile Edge Computing architectures considered in 5G. 2. Context The “network side” of caching: While caching had been widely studied in peer-to-peer systems and content distribution networks (CDNs), the seminal work of femto-caching was among the first to propose caching content on a set of inexpensive edge nodes (called ``helper’’) nodes, which can be, for example, femto-cells with limited backhaul capacity. In this context, the network topology and the content popularity distribution (both assumed to be known) are used to solve the optimal content placement problem: which contents to store in which cache. A number of follow-up works within the ”femto-caching” framework have appeared, considering aspects such as storage in user devices, constraints on transmission capacity per helper leading to a joint placement and routing problem, multi-layer video streaming where video quality can be traded off with hit rate, and multicast through multiple helper nodes, using LTE’s eMBMS framework. A number of other aspects have also been considered in this framework, such as local popularity patterns, using social relations of users to improve prefetching, as well as cache replacement policies. These latter works use stochastic geometry methods to model the random placement of users and helper nodes and resulting topology, rather than assuming a given one. The common denominators between these works can be summarized as follows: (i) The main bottleneck is the backhaul link (ii) the transmission phase is ignored or simplified, (iii) global caching gains stem from coverage overlaps between nearby cells. The “communications side” of caching: When multiple nearby users are requesting content at the same time, the content delivery side of the problem in a wireless setup becomes just as important as the placement problem. Recent work by Maddah-Ali and Niesen revealed quite interesting findings about the fundamental gains achievable by jointly considering caching and coded transmissions. In a simple setup with a catalog of N files, K receiving users each with a request for any one of the N files, storage at the user devices with a cache of size of M (< N) files, the authors show that the amount of time T can be reduced not only due to the local caching gain (M/N) of traditional caching schemes, but also due to a global caching gain that scales with the total size of caching memory, KM. While the above results were derived in a rather stylized setting, a number of follow-up studies have tried to address various shortcomings of the original work, towards distributed implementations, non-uniform popularity distributions, replacement algorithms, hierarchical topologies, etc. Beyond these works, caching has also been considered in the context of advanced cooperative transmission techniques on the physical layer, such as CoMP. CoMP can significantly improve performance, but requires all BS involved to exchange contents, to create a virtual MIMO channel, a large overhead on the already taxed backhaul. However, if every BS involved already stores the contents requested by every user involved, then a collaborative transmission can be performed by just knowing the channel matrix. Some authors argue that the caching and transmission algorithms at each involved BS must be jointly designed in order to facilitate such CoMP opportunities. Finally, in a very recent work of ideas from coded caching are also used to derive fundamental performance bounds on the impact of caching for

Doctorant.e: Giannakas Theodoros