Network slicing has been proposed as a paradigm for 5G+ networks. The operators slice physical resources from the edge all the way to the datacenter, and are responsible to micro-manage the allocation of these resources among tenants bound by predefined Service Level Agreements (SLAs). A key task, for which recent works have advocated the use of Deep Neural Networks (DNNs), is tracking the tenant demand and scaling its resources. Nevertheless, for the edge resources (e.g., RAN), a question arises on whether operators can: (a) scale them fast enough (often in the order of ms) and (b) afford to transmit huge amounts of data towards a remote cloud where such a DNN model might operate. We propose a Distributed DNN (DDNN) architecture for a class of such problems: a small subset of the DNN layers at the edge attempt to act as fast, standalone resource allocator; this is complemented by a mechanism to intelligently offload a percentage of (harder) decisions to additional DNN layers running at a remote cloud. To implement the offloading, we propose: (i) a Bayes-inspired method, using dropout during inference, to estimate the confidence in the local prediction; (ii) a learnable function which automatically classifies samples as “remote” (to be offloaded) or “local”. Using the public Milano dataset, we investigate how such a DDNN should be trained and operated to address (a) and (b). In some cases, our offloading methods are near-optimal, resolving up to 50% of decisions locally with little or no penalty on the allocation cost.

To accommodate low latency and computation-intensive services, such as the Internet-of-Things (IoT), 5G networks are expected to have cloud and edge computing capabilities. To this end, we consider a generic network setup where devices, performing analytics-related tasks, can partially process a task and offload its remainder to base stations, which can then reroute it to cloud and/or to edge servers. To account for the potentially unpredictable traffic demands and edge network dynamics, we formulate the resource allocation as an online convex optimization problem with service violation constraints and allow limited communication between neighboring nodes. To address the problem, we propose an online distributed (across the nodes) primal-dual algorithm and prove that it achieves sublinear regret and violation; in fact, the achieved bound is of the same order as the best known centralized alternative. Our results are further supported using the publicly available Milano dataset.

Caching algorithms try to predict content popularity, and place the content closer to the users. Additionally, nowadays requests are increasingly driven by recommendation systems (RS). These important trends, point to the following: \emph{make RSs favor locally cached content}, this way operators reduce network costs, and users get better streaming rates. Nevertheless, this process should preserve the quality of the recommendations (QoR). In this work, we propose a Markov Chain model for a stochastic, recommendation-driven \emph{sequence} of requests, and formulate the problem of selecting high quality recommendations that minimize the network cost \emph{in the long run}. While the original optimization problem is non-convex, it can be convexified through a series of transformations. Moreover, we extend our framework for users who show preference in some positions of the recommendations' list. To our best knowledge, this is the first work to provide an optimal polynomial-time algorithm for these problems. Finally, testing our algorithms on real datasets suggests significant potential, e.g.,$2\times$improvement compared to baseline recommendations, and 80\% compared to a greedy network-friendly-RS (which optimizes the cost for I.I.D. requests), while preserving at least 90\% of the original QoR. Finally, we show that taking position preference into account leads to additional performance gains.

Network slicing in beyond 5G systems facilitates the creation of customized virtual networks/services, referred to as “slices”, on top of the physical network infrastructure. Efficient and dynamic orchestration of slices is needed to ensure the stringent and diverse service level agreements (SLAs) required by different services. In this paper, we provide a model that attempts to capture the problem of dynamic slice embedding and reconfiguration supporting a multi-domain setup and diverse, end-to-end SLAs. We then show that such problems can be optimally solved, in theory, with (tabular) Reinforcement Learning algorithms (e.g., Q-learning) even under, a priori, unknown demand dynamics for each slice. Nevertheless, the state and action complexity of such algorithms is prohibitive, even for very small scenarios. To this end, we propose a novel scheme based on independent DQN agents: The DQN component implements approximate Q-learning, based on simple, generic DNNs for value function approximation, radically reducing state space complexity; the independent agents then tackle the equally important issue of exploding action complexity arising from the combinatorial nature of embedding multiple VNFs per slice, multiple slices, over multiple domains and computing nodes therein. Using realistic data, we show that the proposed algorithm reduces convergence time by orders of magnitude with minimum penalty of decision optimality.

Network slicing has been proposed as a paradigm for 5G+ networks. The operators slice physical resources from the edge, all the way to datacenter, and are responsible to micromanage the allocation of these resources among tenants bound by predefined Service Level Agreements (SLAs). A key task, for which recent works have advocated the use of Deep Neural Networks (DNNs), is tracking the tenant demand and scaling its resources. Nevertheless, for edge resources (e.g. RAN), a question arises whether operators can: (a) scale edge resources fast enough (often in the order of ms) and (b) afford to transmit huge amounts of data towards a cloud where such a DNN-based algorithm might operate. We propose a Distributed-DNN architecture for a class of such problems: a small subset of the DNN layers at the edge attempt to act as fast, standalone resource allocator; this is coupled with a Bayesian mechanism to intelligently offload a subset of (harder) decisions to additional DNN layers running at a remote cloud. Using the publicly available Milano dataset, we investigate how such a DDNN should be jointly trained, as well as operated, to efficiently address (a) and (b), resolving up to 60% of allocation decisions locally with little or no penalty on the allocation cost.

Caching content over CDNs or at the network edge has been solidified as a means to improve network cost and offer better streaming experience to users. Furthermore, nudging the users towards low-cost content has recently gained momentum as a strategy to boost network performance. We focus on the problem of optimal policy design for Network Friendly Recommendations (NFR). We depart from recent modeling attempts, and propose a Markov Decision Process (MDP) formulation. MDPs offer a unified framework that can model a user with random session length. As it turns out, many state-of-the-art approaches can be cast as subcases of our MDP formulation. Moreover, the approach offers flexibility to model users who are reactive to the quality of the received recommendations. In terms of performance, for users consuming an arbitrary number of contents in sequence, we show theoretically and using extensive validation over real traces that the MDP approach outperforms myopic algorithms both in session cost as well as in offered recommendation quality. Finally, even compared to optimal state-of-art algorithms targeting specific subcases, our MDP framework is significantly more efficient, speeding the execution time by a factor of 10, and enjoying better scaling with the content catalog and recommendation batch sizes.

As mobile traffic is dominated by content services (e.g., video), which typically use recommendation systems, the paradigm of network-friendly recommendations (NFR) has been proposed recently to boost the network performance by promoting content that can be efficiently delivered (e.g., cached at the edge). NFR increase the network performance, however, at the cost of being unfair towards certain contents when compared to the standard recommendations. This unfairness is a side effect of NFR that has not been studied in literature. Nevertheless, retaining fairness among contents is a key operational requirement for content providers. This paper is the first to study the fairness in NFR, and design fair-NFR. Specifically, we use a set of metrics that capture different notions of fairness, and study the unfairness created by existing NFR schemes. Our analysis reveals that NFR can be significantly unfair. We identify an inherent trade-off between the network gains achieved by NFR and the resulting unfairness, and derive bounds for this trade-off. We show that existing NFR schemes frequently operate far from the bounds, i.e., there is room for improvement. To this end, we formulate the design of Fair-NFR (i.e., NFR with fairness guarantees compared to the baseline recommendations) as a linear optimization problem. Our results show that the Fair-NFR can achieve high network gains (similar to non-fair-NFR) with little unfairness.

Caching popular content at the network edge can benefit both the operator and the client by alleviating the backhaul traffic and reducing access latency, respectively. Recommendation systems, on the other hand, try to offer interesting content to the user and impact her requests, but independently of the caching policy. Nevertheless, it has been recently proposed that designing caching and recommendation policies separately is suboptimal. Caching could benefit by knowing the recommender's actions in advance, and recommendation algorithms could try to favor cached content (among equally interesting options) to improve network performance and user experience. In this paper we tackle the problem of optimally making caching and recommendation decisions jointly, in the context of the recently introduced “soft cache hits” setup. We show that even the simplest (one user, one cache) problem is NP-hard, but that the most generic problem (multiple users, femtocaching network) is approximable to a constant. To the best of our knowledge, this is the first polynomial algorithm with approximation guarantees for the joint problem. Finally, we compare our algorithm to existing schemes using a range of real-world data-sets.

Caching has recently attracted a lot of attention in the wireless communications community, as a means to cope with the increasing number of users consuming web content from mobile devices. Caching offers an opportunity for a win-win scenario: nearby content can improve the video streaming experience for the user, and free up valuable network resources for the operator. At the same time, recent works have shown that recommendations of popular content apps are responsible for a significant percentage of users requests. As a result, some very recent works have considered how to nudge recommendations to facilitate the network (e.g., increase cache hit rates). In this paper, we follow up on this line of work, and consider the problem of designing cache friendly recommendations for long viewing sessions; specifically, we attempt to answer two open questions in this context: (i) given that recommendation position affects user click rates, what is the impact on the performance of such network-friendly recommender solutions? (ii) can the resulting optimization problems be solved efficiently, when considering both sequences of dependent accesses (e.g., YouTube) and position preference? To this end, we propose a stochastic model that incorporates position-aware recommendations into a Markovian traversal model of the content catalog, and derive the average cost of a user session using absorbing Markov chain theory. We then formulate the optimization problem, and after a careful sequence of equivalent transformations show that it has a linear program equivalent and thus can be solved efficiently. Finally, we use a range of real datasets we collected to investigate the impact of position preference in recommendations on the proposed optimal algorithm. Our results suggest more than 30% improvement with respect to state-of-the-art methods.

Caching has been successfully applied in wired networks, in the context of Content Distribution Networks (CDNs), and is quickly gaining ground for wireless systems. Storing popular content at the edge of the network (e.g, at small cells) is seen as a “win-win” for both the user (reduced access latency) and the operator (reduced load on the transport network and core servers). Nevertheless, the much smaller size of such edge caches, and the volatility of user preferences suggest that standard caching methods do not suffice in this context. What is more, simple popularity-based models commonly used (e.g, IRM) are becoming outdated, as users often consume multiple contents in sequence (e.g. YouTube, Spotify), and this consumption is driven by recommendation systems. The latter presents a great opportunity to bias the recommender to minimize content access cost (e.g, maximizing cache hit rates). To this end, in this paper we first propose a Markovian model for recommendation-driven user requests. We then formulate the problem of biasing the recommendation algorithm to minimize access cost, while maintaining acceptable recommendation quality. We show that the problem is non-convex, and propose an iterative ADMM-based algorithm that outperforms existing schemes, and shows significant potential for performance improvement on real content datasets.

Pushing popular content to small cells with local storage (“helper” nodes) has been proposed to cope with the ever-growing data demand. Nevertheless, the collective storage of a few nearby helper nodes may not suffice to achieve a high hit rate in practice. In this paper, we introduce the concept of “soft cache hits” (SCHs). An SCH occurs if a user's requested content is not in the local cache, but the user can be (partially) satisfied by a related content that is. In case of a cache miss, an application proxy (e.g., YouTube) running close to the helper node (e.g., at a multi-access edge computing server) can recommend the most related files that are locally cached. This system could be activated during periods of predicted congestion, or for selected users (e.g., low-cost plans), to improve cache hit ratio with limited (and tunable) user quality of experience performance impact. Beyond introducing a model for soft cache hits, our next contribution is to show that the optimal caching policy should be revisited when SCHs are allowed. In fact, we show that optimal caching with SCH is NP-hard even for a single cache. To this end, we formulate the optimal femto-caching problem with SCH in a sufficiently generic setup and propose efficient algorithms with provable performance. Finally, we use a large range of real datasets to corroborate our proposal.

Pushing popular content to cheap ``helper'' nodes (e.g., small cells with local storage) during off-peak hours has recently been proposed to cope with the increase in mobile data traffic. If the requested content is available locally at a helper node, both user and operator performance could benefit. Nevertheless, the collective storage of a few nearby helper nodes does not usually suffice to achieve a high hit rate in practice. In this paper, we investigate the concept of ``soft cache hits'' where, if the original content is not available, some locally cached related contents can be recommended. Given that Internet content consumption is entertainment-oriented, we argue that there exist scenarios where a user might accept an alternative content (e.g., better download rate for alternative content, low rate plans), thus avoiding to access expensive/congested links. We formulate the problem of optimal edge caching with soft cache hits in a sufficiently generic setup, propose an efficient algorithm, and analyze the expected gains. We then show using synthetic and real datasets of related video contents that promising caching gains could be achieved in practice.