Search

Neural networks have achieved state of the art performance across a wide variety of machine learning tasks, often with large and computation-heavy models. Inducing sparseness as a way to reduce the memory and computation footprint of these models has seen significant research attention in recent years. In this paper, we present a new method for dynamic sparseness, whereby part of the computations are omitted dynamically, based on the input. For efficiency, we combined the idea of dynamic sparseness with block-wise matrix-vector multiplications. In contrast to static sparseness, which permanently zeroes out selected positions in weight matrices, our method preserves the full network capabilities by potentially accessing any trained weights. Yet, matrix vector multiplications are accelerated by omitting a pre-defined fraction of weight blocks from the matrix, based on the input. Experimental results on the task of language modeling, using recurrent and quasi-recurrent models, show that the proposed method can outperform static sparseness baselines. In addition, our method can reach similar language modeling perplexities as the dense baseline, at half the computational cost at inference time.

In simulating renewable energy systems (RES), researchers tend to focus on the modeling methodology, but less on the impact of data on the obtained results and associated conclusions. Yet, results may be particularly dependent on the data resolution. RES studies on that impact of data resolution mostly consider households with PV systems, leaving industrial sites and wind power systems out of scope. In this paper, we consider a system comprising a medium-sized wind turbine, a high power battery and an industrial load, connected to a distribution grid. Wind and load power were measured concurrently in real time at 1 s resolution for 2 months in summer 2017, at two locations in Belgium. We performed two simulations, one using the real time data at 1 s resolution and the other using the same power data averaged over 10 min intervals. We compared both simulations to calculate three metrics: total self-sufficiency error, battery utilization error and instantaneous self-sufficiency error. We repeated the analysis using different settings in terms of battery capacity, battery C rate limit and load ratio. We conclude that: (i) total self-sufficiency, in all cases, is overestimated, ranging between 0.06 and 3.6%, with errors decreasing for higher battery capacity, (ii) battery utilization is always underestimated with errors ranging between 7.5 and 44.7%, primarily depending on the C rate limit, and (iii) there is a clear correlation between instantaneous self-sufficiency error and instantaneous ratio of the averaged load and wind power, with errors exceeding 100% when the ratio ranges in 0.5–2.

Renewable energy is expected to constitute a significant proportion of electricity production. Further, the global population is increasingly concentrated in cities. We investigate whether it is possible to cost-effectively employ 100% renewable energy sources (RES)-including battery energy storage systems (BESS)-for producing electricity to meet cities' loads. We further analyze the potential to use only RES to meet partial loads, e.g., by meeting load demands only for certain fractions of the time. We present a novel flexible-load methodology and investigate the cost reduction achieved by shifting fractions of load across time. We use it to evaluate the impacts of exploiting flexibility on making a 100% RES scenario cost effective. For instance, in a case study for Kortrijk, a typical Belgian city with around 75,000 inhabitants, we find that from a purely economic viewpoint, RES-BESS systems are not cost effective even with flexible loads: RES-BESS system costs must decrease to around 40% and 7% (around 0.044 (sic)/kWh and 0.038 (sic)/kWh), respectively, of the reference levelized costs of electricity to cost-effectively supply the city's load demand. These results suggest that electricity alone may not lead to high RES penetration, and integration between electricity, heat, transport, and other sectors is crucial. (C) 2018 Elsevier Ltd. All rights reserved.

This paper presents a data-driven modeling approach for developing control-oriented thermal models of buildings. These models are developed with the objective of reducing energy consumption costs while controlling the indoor temperature of the building within required comfort limits. To combine the interpretability of white/gray box physics models and the expressive power of neural networks, we propose a physics informed neural network approach for this modeling task. Along with measured data and building parameters, we encode the neural networks with the underlying physics that governs the thermal behavior of these buildings. Thus, realizing a model that is guided by physics, aids in modeling the temporal evolution of room temperature and power consumption as well as the hidden state, i.e., the temperature of building thermal mass for subsequent time steps. The main research contributions of this work are: (1) we propose two variants of physics informed neural network architectures for the task of control-oriented thermal modeling of buildings, (2) we show that training these architectures is data-efficient, requiring less training data compared to conventional, non-physics informed neural networks, and (3) we show that these architectures achieve more accurate predictions than conventional neural networks for longer prediction horizons. We test the prediction performance of the proposed architectures using simulated and real-word data to demonstrate (2) and (3) and show that the proposed physics informed neural network architectures can be used for this control-oriented modeling problem., 14 pages, 7 figures

Non-intrusive load monitoring methods aim to disaggregate the total power consumption of a household into individual appliances by analyzing changes in the voltage and current measured at the grid connection point of the household. The goal is to identify the active appliances, based on their unique fingerprint. Most state-of-the-art classification algorithms rely on the assumption that all events in the data stream are triggered by known appliances, which is often not the case. This paper proposes a method capable of detecting previously unidentified appliances in an automated way. For this, appliances represented by their VI trajectory are mapped to a newly learned feature space created by a siamese neural network such that samples of the same appliance form tight clusters. Then, clustering is performed by DBSCAN allowing the method to assign appliance samples to clusters or label them as 'unidentified'. Benchmarking on PLAID and WHITED shows that an F-1.macro-measure of respectively 0.90 and 0.85 can be obtained for classifying the unidentified appliances as 'unidentified'.

So far, the biggest photovoltaic park in Belgium has been injecting all its energy into the electric distribution grid through a power purchase agreement with an electricity supplier. Due to decreasing and volatile wholesale electricity prices, the industrial partners/owners of the photovoltaic park are considering hydrogen storage in an attempt to increase the value proposition of their renewable energy installation. A major objective of the present work is to show how the utilization factor of the electrolyzer is affected by the design of the power supply system when the latter consists only of renewable energy sources instead of using the electric grid. Different hybrid designs were developed, by combining the existing photovoltaic source with wind power and state-of-the-art energy storage technologies (Vanadium Redox Flow or Lithium NMC). Finally, four scenarios were investigated, all considering a 1 MW PEM electrolyzer: A) 15 MW PV, B) 15 MW PV, 2MW Wind, C) 15 MW PV, 2 MW Wind, Battery, D) 15 MW PV, 15 MW Wind. The utilization factor was found as follows, for each scenario respectively: A) 41,5%, B) 65,5%, C) 66,0–86,0%, D) 82,0%. Furthermore, the analysis was extended to include economic evaluations (i.e. payback period, accumulated profit), specifically concerning scenario B and C. The results of this study lead to a number of conclusions such as: i) The utilization of the electrolyzer is limited when its power supply is intermittent. ii) Compared to PV, wind power makes larger contribution to the increase of the utilization factor, iii) 100% utilization can be achieved only if an energy storage system co-exists. iv) With a utilization factor at 65,5% scenario B can deliver a payback period in less than 8 years, if hydrogen is sold above 5€/kg. An analytic overview of all conclusions is presented in the last section of the paper.

The electric vehicle (EV) flexibility, indicates to what extent the charging load can be coordinated (i.e., to flatten the load curve or to utilize renewable energy resources). However, such flexibility is neither well analyzed nor effectively quantified in literature. In this paper we fill this gap and offer an extensive analysis of the flexibility characteristics of 390k EV charging sessions and propose measures to quantize their flexibility exploitation. Our contributions include: (1) characterization of the EV charging behavior by clustering the arrival and departure time combinations that leads to the identification of type of EV charging behavior, (2) in-depth analysis of the characteristics of the charging sessions in each behavioral cluster and investigation of the influence of weekdays and seasonal changes on those characteristics including arrival, sojourn and idle times, and (3) proposing measures and an algorithm to quantitatively analyze how much flexibility (in terms of duration and amount) is used at various times of a day, for two representative scenarios. Understanding the characteristics of that flexibility (e.g., amount, time and duration of availability) and when it is used (in terms of both duration and amount) helps to develop more realistic price and incentive schemes in DR algorithms to efficiently exploit the offered flexibility or to estimate when to stimulate additional flexibility. (C) 2017 Elsevier Ltd. All rights reserved.

Since the inception of non-intrusive appliance load monitoring (NILM), extensive research has focused on identifying an effective set of features that allows to form a unique appliance signature to discriminate various loads. Although an abundance of features are reported in literature, most works use only a limited subset of them. A systematic comparison and combination of the available features in terms of their effectiveness is still missing. This paper, as its first contribution, offers a concise and updated review of the features reported in literature for the purpose of load identification. As a second contribution, a systematic feature elimination process is proposed to identify the most effective feature set. The analysis is validated on a large benchmark dataset and shows that the proposed feature elimination process improves the appliance classification accuracy for all the appliances in the dataset compared to using all the features or randomly chosen subsets of features. (C) 2017 Elsevier B.V. All rights reserved.

A basic but crucial step to increase efficiency and save energy in residential settings is to have an accurate view of energy consumption. To monitor residential energy consumption cost-effectively, i.e., without relying on per-device monitoring equipment, non-intrusive load monitoring (NILM) provides an elegant solution. The aim of NILM is to disaggregate the total power consumption (as measured, e.g., by smart meters at the grid connection point of the household) into individual devices' power consumption, using machine learning techniques. An essential building block of NILM is event detection: detecting when appliances are switched on or off. Current state-of-the-art methods face two open issues. First, they are typically not robust to differences in base load power consumption and secondly, they require extensive parameter optimization. In this paper, both problems are addressed. First two novel and robust algorithms are proposed: a modified version of the chi-squared goodness-of-fit (x(2) GOF) test and an event detection method based on cepstrum smoothing. Then, a workflow using surrogate-based optimization (SBO) to efficiently tune these methods is introduced. Benchmarking on the BLUED dataset shows that both suggested algorithms outperform the standard x2 GOF test for traces with a higher base load and that they can be optimized efficiently using SBO. (C) 2017 Elsevier B.V. All rights reserved.

Smart grids collect large volumes of smart meter data in the form of time series, or so-called load patterns. We outline the applications that benefit from analyzing this data (ranging from customer segmentation to operational system planning), and propose two-stage load pattern clustering. The first stage is performed per individual user and identifies the various typical daily power usage patterns (s)he exhibits. The second stage takes those typical user patterns as input to group users that are similar. To improve scalability, we use fast wavelet transformation (FWT) of the time series data, which reduces the dimensionality of the feature space where the clustering algorithm operates (i.e., from ${N}$ data points in the time domain to log ${N}$ ). Another qualitative benefit of FWT is that patterns that are identical in shape, but just differ in a (typically small) time shift still end up in the same cluster. Furthermore, we use ${g}$ -means instead of ${k}$ -means as the clustering algorithm. Our comprehensive set of experiments analyzes the impact of using FWT versus time-domain features, and ${g}$ - versus ${k}$ -means, to conclude that in terms of cluster quality metrics our system is comparable to state-of-the-art methods, while being more scalable (because of the dimensionality reduction).

Evaluation of search engines relies on assessments of search results for selected test queries, from which we would ideally like to draw conclusions in terms of relevance of the results for general (e.g., future, unknown) users. In practice however, most evaluation scenarios only allow us to conclusively determine the relevance towards the particular assessor that provided the judgments. A factor that cannot be ignored when extending conclusions made from assessors towards users, is the possible disagreement on relevance, assuming that a single gold truth label does not exist. This paper presents and analyzes the Predicted Relevance Model (PRM), which allows predicting a particular result's relevance for a random user, based on an observed assessment and knowledge on the average disagreement between assessors. With the PRM, existing evaluation metrics designed to measure binary assessor relevance, can be transformed into more robust and effectively graded measures that evaluate relevance towards a random user. It also leads to a principled way of quantifying multiple graded or categorical relevance levels for use as gains in established graded relevance measures, such as normalized discounted cumulative gain (nDCG), which nowadays often use heuristic and data-independent gain values. Given a set of test topics with graded relevance judgments, the PRM allows evaluating systems on different scenarios, such as their capability of retrieving top results, or how well they are able to filter out non-relevant ones. Its use in actual evaluation scenarios is illustrated on several information retrieval test collections., Accepted for publication in Springer Information Retrieval Journal, special issue on Information Retrieval Evaluation using Test Collections

We address the dimensioning of infrastructure, comprising both network and server resources, for large-scale decentralized distributed systems such as grids or clouds. We design the resulting grid/cloud to be resilient against network link or server failures. To this end, we exploit relocation: Under failure conditions, a grid job or cloud virtual machine may be served at an alternate destination (i.e., different from the one under failure-free conditions). We thus consider grid/cloud requests to have a known origin, but assume a degree of freedom as to where they end up being served, which is the case for grid applications of the bag-of-tasks (BoT) type or hosted virtual machines in the cloud case. We present a generic methodology based on integer linear programming (ILP) that: 1) chooses a given number of sites in a given network topology where to install server infrastructure; and 2) determines the amount of both network and server capacity to cater for both the failure-free scenario and failures of links or nodes. For the latter, we consider either failure-independent (FID) or failure-dependent (FD) recovery. Case studies on European-scale networks show that relocation allows considerable reduction of the total amount of network and server resources, especially in sparse topologies and for higher numbers of server sites. Adopting a failure-dependent backup routing strategy does lead to lower resource dimensions, but only when we adopt relocation (especially for a high number of server sites): Without exploiting relocation, potential savings of FD versus FID are not meaningful.

The widespread availability of cloud infrastructures is fueling new interest in the thin client computing paradigm. However, current thin client protocols are not designed to handle new content types as often encountered in state-of-the-art applications (e.g. multimedia editing, gaming, multimedia playback). Conveying this content using traditional thin client protocols typically results in a combination of excessive resource usage and low visual quality. In this paper, we propose an approach where the content type can vary for different portions of the screen (e.g. combination of static text and video). Once the different content types have been detected, each of them can be encoded using the most appropriate algorithm. We present two algorithms for this runtime detection. The first algorithm is operating at the pixel level, thereby being independent of the actual thin client protocol used. The second algorithm assumes the presence of a rectangle-based thin client protocol (such as the popular VNC protocol), trading independence for improved performance. The appropriate parameter settings for these algorithms are experimentally determined. Furthermore, their influence is studied in detail in terms of detection accuracy, and the time to perform the algorithms is analysed. Accurate hints are derived within less than 10ms, indicating the high potential of this approach for use in next generation thin client systems.

The evolution toward grid and cloud computing as observed for over a decennium illustrates the crucial role played by (optical) networks in supporting today's applications. In this paper, we start from an overview of the challenging applications in both academic (further referred to as scientific), enterprise (business) and nonprofessional user (consumer) domains. They pose novel challenges, calling for efficient interworking of IT resources, for both processing and storage, as well as the network that interconnects them and provides access to their users. We outline those novel applications' requirements, including sheer performance attributes (which will determine the quality as perceived by end users of the cloud applications), as well as the ability to adapt to changing demands (usually referred to as elasticity) and possible failures (i.e., resilience). In outlining the foundational concepts that provide the building blocks for grid/cloud solutions that meet the stringent application requirements we highlight, a prominent role is played by optical networking. The pieces of the solution studied in this respect span the optical transport layer as well as mechanisms located in higher layers (e.g., anycast routing, virtualization) and their interworking (e.g., through appropriate control plane extensions and middleware). Based on this study, we conclude by identifying challenges and research opportunities that can enable future-proof optical cloud systems (e.g., pushing the virtualization paradigms to optical networks).

Grids use a form of distributed computing to tackle complex computational and data processing problems scientists are presented with today. When designing an (optical) network supporting grids, it is essential that it can overcome single network failures, for which several protection schemes have been devised in the past. In this work, we extend the existing Shared Path protection scheme by incorporating the anycast principle typical of grids: a user typically does not care on what specific server this job gets executed and is merely interested in its timely delivery of results. Therefore, in contrast with Classical Shared Path protection (CSP), we will not necessarily provide a backup path between the source and the original destination. Instead, we allow to relocate the job to another server location if we can thus provide a backup path which comprises less wavelengths than the one CSP would suggest. We assess the bandwidth savings enabled by relocation in a quantitative dimensioning case study on an European and an American network topology, exhibiting substantial savings of the number of required wavelengths (in the order of 11-50%, depending on network topology and server locations). We also investigate how relocation affects the computational load on the execution servers. The case study is based on solving a grid network dimensioning problem: we present Integer Linear Programming (ILP) formulations for both the traditional CSP and the new resilience scheme exploiting relocation (SPR). We also outline a strategy to deal with the anycast principle: assuming we are given just the origins and intensity of job arrivals, we derive a static (source,destination)-based demand matrix. The latter is then used as input to solve the network dimensioning ILP for an optical circuit-switched WDM network.

Support for anycast in the IP network layer allows one source node to contact a single member out of a group of destination nodes configured with the same IP address. Due to the stateless nature of the IP protocol, subsequent packets from the same source node targeted at the same anycast group may arrive at different group members. Consequently, native IP anycast cannot be applied directly to support distributed session-based services. For this reason, an anycast overlay architecture combining the transparency offered by native anycast with support for stateful communications has been proposed. In this paper, we investigate the operational impact of deploying this overlay architecture. Performance evaluation of a data plane prototype implementation for an anycast overlay node shows that high throughput and small latency can be achieved. Additionally, we show how threshold-based update triggering, in combination with an appropriate inter-proxy update strategy, delivers control plane accuracy with minimal network overhead.

A key problem in Grid networks is how to efficiently manage the available infrastructure, in order to satisfy user requirements and maximize resource utilization. This is in large part influenced by the algorithms responsible for the routing of data and the scheduling of tasks. In this paper, we present several multi-cost algorithms for the joint scheduling of the communication and computation resources that will be used by a Grid task. We propose a multi-cost scheme of polynomial complexity that performs immediate reservations and selects the computation resource to execute the task and determines the path to route the input data. Furthermore, we introduce multi-cost algorithms that perform advance reservations and thus also find the starting times for the data transmission and the task execution. We initially present an optimal scheme of non-polynomial complexity and by appropriately pruning the set of candidate paths we also give a heuristic algorithm of polynomial complexity. Our performance results indicate that in a Grid network in which tasks are either CPU- or data-intensive (or both), it is beneficial for the scheduling algorithm to jointly consider the computational and communication problems. A comparison between immediate and advance reservation schemes shows the trade-offs with respect to task blocking probability, end-to-end delay and the complexity of the algorithms.

Recent years have demonstrated the limited scalability of electronic switching to realize transport networks. In response, all-optical switching has been identified as a candidate solution to enable high-capacity networking in the future. One of the fundamental challenges is to efficiently support a wide range of traffic patterns, and thus emerges the need for equipment that is both practical and economical to construct and deploy. We have previously proposed the use of multi-granular optical cross-connects (MG-OXC), which support switching on both the wavelength and sub-wavelength level. To this end, the MG-OXCs are equipped with cheap, highly scalable slow switching fabrics, as well as a small number of expensive fast switching ports. The goal of this work is two-fold: first to demonstrate that a small number of fast switching ports suffices to support a wide range of traffic requirements, and second that multi-granular optical switching can offer cost-benefits on a network-wide scale. The first objective is studied through simulation analysis of a single switching node, and results indicate that a limited number of fast switching ports can significantly improve burst blocking performance over slow only switches. Furthermore, under certain circumstances, the MG-OXC can even approach the performance of a fast only switch design. Secondly, we introduce an Integer Linear Programming model for the total network installation cost, and our evaluation indicates that multi-granular optical switching can be a cost-effective solution on the network level, in comparison to slow only or fast only approaches. Furthermore, we can achieve reduced costs of individual OXC nodes, which allows us to minimize scalability problems corresponding to emerging fast switching fabrics.

Optical packet switching (OPS) has been proposed as a strong candidate for future metro networks. This paper assesses the viability of an OPS-based ring architecture as proposed within the research project DAVID (Data And Voice Integration on DWDM), funded by the European Commission through the Information Society Technologies (IST) framework. Its feasibility is discussed from a physical-layer point of view, and its limitations in size are explored. Through dimensioning studies, we show that the proposed OPS architecture is competitive with respect to alternative metropolitan area network (MAN) approaches, including synchronous digital hierarchy, resilient packet rings (RPR), and star-based Ethernet. Finally, the proposed OPS architectures are discussed from a logical performance point of view, and a high-quality scheduling algorithm to control the packet-switching operations in the rings is explained.

We address the problem of congestion resolution in optical packet switching (OPS). We consider a fairly generic all-optical packet switch architecture with a feedback optical buffer constituted of fiber delay lines (FDL). Two alternatives of switching granularity are addressed for a switch operating in a slotted transfer mode: switching at the slot level (i.e., fixed length packets of a single slot) or at the burst level (variable length packets that are integer multiples of the slot length). For both cases, we show that in spite of the limited queuing resources, acceptable performance in terms of packet loss can be achieved for reasonable hardware resources with an appropriate design of the time/wavelength scheduling algorithms. Depending on the switching units (slots or bursts), an adapted scheduling algorithm needs to be deployed to exploit the bandwidth and buffer resources most efficiently.

The major goal of optical packet switching (OPS) is to match switching technology to the huge capacities provided by (D)WDM. A crucial issue in packet switched networks is the avoidance of packet losses stemming from contention. In OPS, contention can be solved using a combination of exploitation of the wavelength domain (through wavelength conversion) and buffering. To provide optical buffering, fiber delay lines (FDLs) are used. In this paper, we focus on an optical packet switch with recirculating FDL buffers and wavelength converters. We introduce the Markovian arrival process with marked transitions (MMAP), which has very desirable properties as a traffic model for OPS performance assessment. Using this model, we determine lower and upper bounds for the packet loss rate (PLR) achieved by the aforementioned switch. The calculation of the PLR bounds through matrix analytical methods is repeated for a wide range of traffic conditions, including highly non-uniform traffic, both in space (i.e., packet destinations) and time (bursty traffic). The quality of these bounds is verified through comparison with simulation results.

The major goal of optical packet switching (OPS) is to match switching technology to the huge capacities provided by (D)WDM. We study optical packet switches with recirculating fiber delay line (FDL) buffers. Through simulation, we have assessed the logical performance of a single optical packet router (OPR), focusing on packet loss rate (PLR). By verifying that our scheduling algorithm does not alter the traffic profile characteristics from in- to output, we illustrate how the single node results can be used to assess network-wide performance. We use the capability of assessing end-to-end PLRs to develop network-wide routing algorithms designed to minimize the maximal PLR occurring in the network. In case studies on pan-European networks, we first compare two algorithm variants and thereafter we compare the PLR-based routing algorithm with both load balancing and shortest path routing. While load balancing achieves PLRs that are multiple orders of magnitude lower than shortest path routing, the PLR-based algorithm can reach PLRs up to two orders of magnitude better. The improvement in PLR comes at the price of only a small increase in used bandwidth (a few percent). Subsequently we show that the discussed PLR-based routing algorithm can be easily extended to multiple priorities. By introducing multiple priorities we can keep the loss rates for high priority traffic very low. However, it may lead to an increase of the obtained minimal max-PLR value for low priority traffic. But as we prove this increase to be limited, the cost of introducing multiple priorities is small.

In this paper, promising technologies and a network architecture are presented for future optical packet switched networks. The overall network concept is presented and the major choices are highlighted and compared with alternative solutions. Both long and shorter term approaches are considered, as well as both the wide-area network and multiple-area networks parts of the network. The results presented in this paper were developed in the frame of the research project DAVID (Data And Voice Integration over DWDM) project, funded by the European Commission through the IST-framework.

The explosive growth of data traffic-for example, due to the popularity of the Internet-poses important emerging network requirements on today's telecommunication networks. This paper describes how core networks will evolve to optical transport networks (OTNs), which are optimized for the transport of data traffic, resulting in an IP-directly-over-OTN paradigm. Special attention is paid to the survivability of such data-centric optical networks. This becomes increasingly crucial since more and more traffic is multiplexed onto a single fiber (e.g., 160/spl times/10 Gb/s), implying that a single cable cut can affect incredible large traffic volumes. In particular, this paper is tackling multilayer survivability problems, since a data-centric optical network consists of at least an IP and optical layer. In practice, this means that the questions "in which layer or layers should survivability be provided?" and "if multiple layers are chosen for this purpose, then how should this functionality in these layers be coordinated?" have to be answered. In addition to a theoretical study, some case studies are presented in order to illustrate the relevance of the described issues and to help in strategic planning decisions. Two case studies are studying the problem from a capacity viewpoint. Another case study presents simulations from a timing/throughput performance viewpoint.

In recent years, telecommunication networks have faced explosive (IP) traffic growth. As traffic keeps growing, network reliability gains more and more importance. This article investigates to which extent switched connections and fast connection provisioning, typical for intelligent optical networks (IONs), can be used to provide resilience in an IP-over-optical multilayer network scenario. This solution, based on transport network flexibility, is compared with more traditional static multilayer resilience schemes in terms of cost (capacity) requirements and operational (dis)advantages.

In cloud-like scenarios, demand is served at one of multiple possible data center (DC) destinations. Usually, the exact DC that is used can be freely chosen, which leads to an anycast routing problem. Furthermore, the demand volume is expected to change over time, e.g., following a diurnal pattern. Given that virtually all application domains today rely heavily on cloud-like services, it is important that the backbone networks connecting users to the DCs are resilient against failures. In this paper, we consider the problem of resiliently routing multi-period traffic: we need to find routes to both a primary DC and a backup DC (to be used in the case of failure of the primary one, or of the network connection to it), and also account for synchronization traffic between the primary and backup DCs. We formulate this as an optimization problem and adopt column generation, using a path formulation in two sub-problems: the (restricted) master problem selects "configurations" to use for each demand in each of the time epochs it lasts, while the pricing problem (PP) constructs a new "configuration" that can lead to lower overall costs (which we express as the number of network resources, i.e., bandwidth, required to serve the demand). Here, a "configuration" is defined by the network paths followed from the demand source to each of the two selected DCs, as well as that of the synchronization traffic in between the DCs. Our decomposition allows for PPs to be solved in parallel, for which we quantitatively explore the reduction in the time required to solve the overall routing problem. The key question that we address with our model is an exploration of the potential benefits of rerouting traffic from one time epoch to the next: we compare several (re) routing strategies, allowing traffic that spans multiple time periods to i) not be rerouted in different periods, ii) only change the backup DC and routes, or iii) freely change both primary and backup DC choices and the routes toward them.

Rising energy costs and climate change have led to an increased concern for energy efficiency (EE). As information and communication technology is responsible for about 4% of total energy consumption worldwide, it is essential to devise policies aimed at reducing it. In this paper, we propose a routing and scheduling algorithm for a cloud architecture that targets minimal total energy consumption by enabling switching off unused network and/or information technology (IT) resources, exploiting the cloud-specific anycast principle. A detailed energy model for the entire cloud infrastructure comprising a wide-area optical network and IT resources is provided. This model is used to make a single-step decision on which IT end points to use for a given request, including the routing of the network connection toward these end points. Our simulations quantitatively assess the EE algorithm's potential energy savings but also assess the influence this may have on traditional quality-of-service parameters such as service blocking. Furthermore, we compare the one-step scheduling with traditional scheduling and routing schemes, which calculate the resource provisioning in a two-step approach (selecting first the destination IT end point and subsequently using unicast routing toward it). We show that depending on the offered infrastructure load, our proposed one-step calculation considerably lowers the total energy consumption (reduction up to 50%) compared to the traditional iterative scheduling and routing, especially in low- to medium-load scenarios, without any significant increase in the service blocking.

Nowadays, the Quality of Service (QoS) in Optical Grids has become a key issue. An important QoS factor is resiliency, namely the ability to survive from certain network failures. Although several traditional network protection schemes were devised in the past, they are not optimized for Optical Grids. In an earlier work, we proposed relocation strategies providing backup paths to alternate destinations, exploiting the anycast routing principle of grids. To show the advantage of relocation (compared to traditional network protection) in terms of reduced network capacity, we formulated the network dimensioning problem as an Integer Linear Program (ILP). Yet, its solution exhibited very poor scalability and appeared not practical for reasonably large scale case studies. Therefore, we propose a novel formulation for the relocation protection scheme, using column generation (CG). This approach decomposes the original ILP into two parts, specifically a Restricted Master Problem (RMP) and a Pricing Problem (PP) which are iteratively and alternatively solved until the optimality condition is satisfied. Such a CG decomposition has a significant impact on the complexity of the model, leading to a significant improvement over previous ILPs in term of scalability and running times. We demonstrate that the CG method is highly scalable and generates nearly optimal solutions using case studies with up to 300 connections, showing it to be highly competitive with a previously proposed heuristic. We also perform some comparisons of the anycast scheme with the classical shared path protection on larger network instances.

In this paper, we address the issue of resiliency against single link network failures in optical grids and show how the anycast routing principle, which is typical of grids, can be exploited in providing efficient shared path protection. We investigate two different integer linear program models for the full anycast routing problem, deciding on the primary and backup server locations as well as on the lightpaths toward them. The first model is a classical integer linear programming (ILP) model, which lacks scalability. The second model is a large-scale optimization model which can be efficiently solved using column generation techniques. We also design two new heuristics: the first one is an improvement of a previously proposed one which, although providing near optimal solutions, lacks scalability, while the second one is highly scalable, at the expense of reduced accuracy. Numerical results are presented for three mesh networks with varying node degrees. They allow an illustration of the scalability of the newly proposed approaches. Apart from highlighting the difference in performance (i.e., scalability and optimality) among the algorithms, our case studies demonstrate the bandwidth savings that can be achieved by exploiting relocation rather than using a backup path to the original (failure-free) destination site. Numerical results for varying network topologies, as well as different numbers of server sites show that relocation allows bandwidth savings in the range of 7-21%.

For traditional optical network dimensioning, a plethora of algorithms exists to design the amount of network resources required to accommodate a given amount of traffic, expressed as a (source, destination)-based traffic matrix. In optical Grid dimensioning, however, the anycast principle applies: Grid users do not really care where exactly their tasks (Grid jobs) end up being executed. Thus, the destination of traffic is not known beforehand and traditional dimensioning algorithms are not applicable. In this paper we propose a mean field calculation method to analytically derive the traffic matrix for given job arrival intensities at the originating Grid sites (the sources). We also indicate how it can be integrated in a stepwise dimensioning approach to compute not only the amount of network resources, but also Grid resources (computational and/or storage). Hence it forms part of a solution for Grid dimensioning: determining how many servers to provide, where to place them, and which network to install for interconnecting server sites and users generating Grid jobs.

A fundamental issue in all-optical switching is to offer efficient and cost-effective transport services for a wide range of bandwidth granularities. This paper presents multi-granular optical cross-connect (MG-OXC) architectures that combine slow (ms regime) and fast (ns regime) switch elements, in order to support optical circuit switching (OCS), optical burst switching (OBS), and even optical packet switching (OPS). The MG-OXC architectures are designed to provide a cost-effective approach, while offering the flexibility and reconfigurability to deal with dynamic requirements of different applications. All proposed MG-OXC designs are analyzed and compared in terms of dimensionality, flexibility/reconfigurability, and scalability. Furthermore, node level simulations are conducted to evaluate the performance of MG-OXCs under different traffic regimes. Finally, the feasibility of the proposed architectures is demonstrated on an application-aware, multi-bit-rate (10 and 40 Gbps), end-to-end OBS testbed.