Edge computing has emerged as a new paradigm to bring cloud applications closer to users for increased performance. Unlike back-end cloud systems which consolidate their resources in a centralized data center location with virtually unlimited capacity, edge-clouds comprise distributed resources at various “computation spots”, each with very limited capacity. In this article, we consider Function-as-a-Service (FaaS) edge-clouds whereapplication providersdeploy their latency-critical functions to process user requests with strict response time deadlines. In this setting, we investigate the problem ofresource provisioningandallocation. After formulating the optimal solution, we propose resource allocation and provisioning algorithms across the spectrum of fully-centralized to fully-decentralized. We evaluate the performance of these algorithms in terms of their ability to utilize CPU resources and meet request deadlines under various system parameters. Our results indicate that practical decentralized strategies, which require no coordination among computation spots, achieve performance that is close to the optimal fully-centralized strategy with coordination overheads.

In a packet network, the routes taken by traffic can be determined according to predefined objectives. Assuming that the network conditions remain static and the defined objectives do not change, mathematical tools such as linear programming could be used to solve this routing problem. However, networks can be dynamic or the routing requirements may change. In that context, Reinforcement Learning (RL), which can learn to adapt in dynamic conditions and offers flexibility of behavior through the reward function, presents as a suitable tool to find good routing strategies. In this work, we train an RL agent, which we call R2L, to address the routing problem. The policy function used in R2L is a neural network and we use an evolution strategy algorithm to determine its weights and biases. We tested R2L in two different scenarios: static and dynamic networks conditions. In the first, we used a 16-node network and experimented with different reward functions, observing that R2L was able to adapt its routing behavior accordingly. Finally, in the second experiment, we used a 5-node network topology where a given link's transmission rate changed during the simulation. In this scenario, we observed that R2L was able to deliver a competitive performance, compared to heuristic benchmarks, with changing network conditions.

It has been claimed that the deployment of fog and edge computing infrastructure is a necessity to make high-performance cloud-based applications a possibility. However, there are a large number of middle-ground latency-sensitive applications such as online gaming, interactive photo editing and multimedia conferencing that require servers deployed closer to users than in globally centralised clouds but do not necessarily need the extreme low-latency provided by a new infrastructure of micro data centres located at the network edge, e.g., in base stations and ISP Points of Presence. In this paper we analyse a snapshot of today's data centres and the distribution of users around the globe and conclude that existing infrastructure provides a sufficiently distributed platform for middle-ground applications requiring a response time of 20-200 ms. However, while placement and selection of edge servers for extreme low-latency applications is a relatively straightforward matter of choosing the closest, providing a high quality of experience for middle-ground latency applications that use the more widespread distribution of today's data centres, as we advocate in this paper, raises new management challenges to develop algorithms for optimising the placement of and the per-request selection between replicated service instances.

In this work, we propose a Deep Learning (DL) based solution to the problem of routing traffic flows in computer networks. Routing decisions can be made in different ways depending on the desired objective and, based on that objective function, optimal solutions can be computed using a variety of techniques, e.g. with mixed integer linear programming. However, determining these solutions requires solving complex optimization problems and, thus, cannot be typically done at runtime. Instead, heuristics for these problems are often created but designing them is non-trivial in many cases. The routing framework proposed here presents an alternative to the design of heuristics, whilst still achieving good performance. This is done by building a DL model trained on the optimal decisions over flows from known traffic demands. To evaluate our solution, we focused on the problem of network congestion, even though a wide range of alternative objectives could be fitted into this framework. We ran experiments using two publicly available datasets of networks with real traffic demands and showed that our solution achieves close-to-optimal network congestion values.

Advanced Internet services increasingly rely on many components to implement their functionality. These composite services have three important features: they are expensive to deploy, components need to be placed intelligently close to the users to improve quality of experience and they will potentially consume significant amounts of bandwidth. This paper presents Triptych, a multi-objective optimisation framework that tries to optimise according these three dimensions to help the three main stakeholders in the Internet ecosystem: users, application providers and network providers. Triptych implements evolutionary computation approaches for this complex problem, which simultaneously optimise service deployment costs, latency-based user utility and network congestion. These algorithms provide possible operating points, bringing important tools for network managements and resource allocation. A large set of simulations under different scenarios are provided to validate the algorithms.

The dominant application in today's Internet is content streaming, which is increasingly relying on caches to meet the stringent conditions on the latency between content servers and end-users. These systems routinely face the challenges of limited bandwidth capacities and network server failures, which degrade caching performance. In this paper, we study the problem of optimally allocating content over a resilient caching network, in which each cache may fail under some situations. Given content request rates and multiple routing paths, we formulate an optimization problem to maximize the expected caching gain, i.e., the reduction of latency due to intermediate caching. The offline version of this problem is NP-hard. We first propose a centralized, offline algorithm and show that a solution with (1-1/e) approximation ratio to the optimal can be constructed. We then propose a distributed ascent algorithm based on the concave relaxation of the expected gain. Informed by the results of our analysis, we finally propose a distributed resilient caching algorithm (DR-Cache) that is simple and adaptive to network failures. We show numerically that DR-Cache significantly outperforms other candidate algorithms under synthetic requests, as well as real world traces over a class of network topologies.

Software-defined Network (SDN) is a new networking paradigm enabling innovation through network programmability. Over past few years, many applications have been built using SDN such as server load balancing, virtual-machine migration, traffic engineering and access control. In this paper, we focus on using SDN for energy-aware routing (EAR). Since traffic load has a small influence on the power consumption of routers, EAR allows putting unused links into sleep mode to save energy. SDN can collect traffic matrix and then computes routing solutions satisfying QoS while being minimal in energy consumption. However, prior works on EAR have assumed that the SDN forwarding table switch can hold an infinite number of rules. In practice, this assumption does not hold since such flow tables are implemented in Ternary Content Addressable Memory (TCAM) which is expensive and power hungry. We consider the use of wildcard rules to compress the forwarding tables. In this paper, we propose optimization methods to minimize energy consumption for a backbone network while respecting capacity constraints on links and rule space constraints on routers. In details, we present two exact formulations using Integer Linear Program (ILP) and introduce efficient heuristic algorithms. Based on simulations on realistic network topologies, we show that using this smart rule space allocation, it is possible to save almost as much power consumption as the classical EAR approach.

The emergence of distributed clouds opens up new research challenges for service deployment. Composite services consist of multiple components, potentially located in different geographical locations, which need to be interconnected and invoked in the correct order according to the overall service work-flow. The placement of composite services over distributed cloud node locations raises new challenges for efficient deployment and management. In this paper, we design exact models of the composite service placement problems using mixed integer linear program, and compare these to solutions based on genetic algorithms. We use a utility function, based initially on latency metrics, to evaluate the quality of service (QoS) of the deployed composite service. By maximizing the utility with respect to deployment cost, our approach can provide good QoS for users while satisfying budget constraints for service providers. Based on simulations using real data-center locations and traffic demand patterns, we show that our algorithms are scalable under a range of scenarios.

This paper is focused on techniques for maximizing utility across all users within a total network transit cost budget. We present a new method for selecting between replicated servers distributed over the Internet. First, we introduce a novel utility framework that factors in quality of service metrics. Then we design an optimization algorithm, solvable in polynomial time, to allocate user requests to servers based on utility while satisfying network transit cost constraints, mapping service names to service instance locators. We then describe an efficient, low overhead distributed model which only requires knowledge of a fraction of the data required by the global optimization formulation. Next, a load-balancing variant of the algorithm is explored that substantially reduces blocking caused by congested servers. Extensive simulations show that our method is scalable and leads to higher user utility compared with mapping user requests to the closest service replica, while meeting network traffic cost constraints. We discuss several options for real-world deployment that require no changes to end-systems based on either the use of SDN controllers or extensions to the current DNS system.

Edge computing has emerged as a new paradigm to bring cloud applications closer to users for increased performance. ISPs have the opportunity to deploy private edge-clouds in their infrastructure to generate additional revenue by providing ultra-low latency applications to local users. We envision a rapid increase in the number of such applications for “edge” networks in the near future with virtual/augmented reality (VR/AR), networked gaming, wearable cognitive assistance, autonomous driving and IoT analytics having already been proposed for edge- clouds instead of the central clouds to improve performance. This raises new challenges as the complexity of the resource allocation problem for multiple services with latency deadlines (i.e., which service to place at which node of the edge-cloud in order to satisfy the latency constraints) becomes significant. In this paper, we propose a set of practical, uncoordinated strategies for service placement in edge-clouds. Through extensive simulations using both synthetic and real-world trace data, we demonstrate that uncoordinated strategies can perform comparatively well with the optimal placement solution, which satisfies the maximum amount of user requests.

Software-defined Networks (SDN), in particular OpenFlow, is a new networking paradigm enabling innovation through network programmability. Over past few years, many applications have been built using SDN such as server load balancing, virtual-machine migration, traffic engineering and access control. In this paper, we focus on using SDN for energy-aware routing (EAR). Since traffic load has a small influence on power consumption of routers, EAR allows to put unused links into sleep mode to save energy. SDN can collect traffic matrix and then computes routing solutions satisfying QoS while being minimal in energy consumption. However, prior works on EAR have assumed that the table of OpenFlow switch can hold an infinite number of rules. In practice, this assumption does not hold since the flow table is implemented with Ternary Content Addressable Memory (TCAM) which is expensive and power-hungry. In this paper, we propose an optimization method to minimize energy consumption for a backbone network while respecting capacity constraints on links and rule space constraints on routers. In details, we present an exact formulation using Integer Linear Program (ILP) and introduce efficient greedy heuristic algorithm. Based on simulations, we show that using this smart rule space allocation, it is possible to save almost as much power consumption as the classical EAR approach.

In this paper we propose bandwidth fair N-Tree algorithm for ALM distribution tree construction and a new protocol for ALM packet replication and distribution, namely Almcast. Both the tree construction algorithm and packet replication/distribution protocol were implemented as proof-of concept by modifying an existing multi-party video conference application. The results show that N-tree algorithm takes less than 3 ms to construct ALM distribution tree for 12 nodes. Almcast implementation enables the intermediate relay node to lookup for next destination, replicate and forward packets as fast as its receiving rate at application layer.