The main motivation for supporting service mobility is the ability to follow end users while providing them with their services with the same look and feel as they move. To do so, a distributed communication paradigm is needed to update the services and a signaling protocol is also needed to track the user’s location and provide multimedia services. The mobile agent paradigm can be used in an elegant way to support service mobility in a mobile environment [1,2]. Mobile agents are entities that can start execution in a node, suspend execution, and move to another node to resume execution. This new paradigm presents many advantages for the realization of service mobility. Those advantages are, but are not limited to, autonomy, mobility and persistency. It is also claimed that mobile agents reduce bandwidth
s c a l a b i l i t y We also measured how the security adaptation latencies depend on system scale, captured by the number of physical domains. Results are shown in Fig- ure 31 . Detection time is slightly increasing with the number of domains. Administrator-
deﬁned OC2 policies become larger with more domains. Thus, more time is needed to match such policies with incoming agent alerts. Policy distribution time is propor- tional to system scale: in our proof-of-concept, the propagation protocol has to be repeated for each device domain. Performance may be improved with a broadcast protocol, instead of the current multi-unicast protocol to distribute security policies. Reaction time is not really impacted: only light distributed veriﬁcations are performed after all reactions to ensure successful cross-domain SLA enforcement. We found la- tency results for combined detection, distribution, and reaction to scale well in terms of domains. However, enforcement times crushed results of all other phases. Thus, further work is required to assess overall architecture scalability. Our current proof-of- concept currently supports 4 domains. Nevertheless, we guess our approach should support far more domains before the framework overhead reaches enforcement costs.
P-Grid [ 2 ] goes one step further than PHT and integrates the indexing algorithm with the underlying routing algorithm. Specifically, P-Grid uses a self-organization process to structure peers directly in a binary trie. In this way, P-Grid dramatically reduces the number of routing hops to answer range queries compared to PHT. In a different approach, Bharambe et al. [ 12 ] propose Mercury a scalable protocol to handle range queries. Mercury differs from previous range-based query systems in that it does not use a cryptographic hash function to index peers and content. Mercury organizes peers in several circular (logical) overlays each dedicated to an attribute of the query space. Then, each peer is statically assigned to a portion of each attribute space and it is responsible of answering the queries that fall in this range. In this way, Mercury eliminates the randomness introduced by the hash function in DHTs and allows to easily handle range queries. The drawback of this approach is that load in Mercury may be unevenly distributed. Thus, explicit overlay modifi- cations are needed to adapt to load variations. This additional cost due to network rewirings can become intolerable in presence of churn, i.e., peer arrivals and depar- tures, and large object sizes.
2 LRI, Univ. Paris-Sud, CNRS, Inria, Universite Paris-Saclay F-91400 Orsay, France
Companies are now using Virtual Reality (VR) for collaborative design reviews on digital mock-ups. These meetings often involve remote collaborators due to current trends towards decentralization of work organization. While lots of previous works proposed dis- tributed architectures for implementing Collaborative Virtual Envi- ronments (CVE), modifying native CAD parts in such environments is challenging. There are two main difficulties: (i) how to directly modify native CAD data (i.e. data used internally in CAD software) from the virtual environment, and (ii) how to manage collaborative modifications of such data by remote users. Most common VR- CAD applications require data conversions before the VR session and post-modifications of original CAD data afterwards. Only a few VR applications allow direct modifications of native CAD data, but they do not support remote collaboration. In this paper, we propose a distributedarchitecture allowing collaborative modifications of native CAD data from remote and heterogeneous platforms. Tech- nically, a VR-CAD server embedding a CAD engine is included in our architecture to load and modify native CAD data according to remote requests. A proof of concept uses our architecture to connect a wall-sized display and a CAVE-like system.
Unfortunately, learning management systems and social media applications are data silos. In other words, data are unavailable on the web. Only people may have access to data, not computers. Reuse and exchange of data among LMS and social tools are only possible by means of API – that is to say manually by mean of one API per tool. On the contrary, semantic web provides a common framework that allows data, information and knowledge to be shared and reused across applications, enterprises, and community boundaries. In such a framework, linked data describes a method of exposing, sharing, and connecting data, information and knowledge on the Web (Bojaars, Breslin et al. 2008; Gruber 2008). It provides a standardized, uniform and generic method for data discovery, distributed queries against several data repositories, integration or semantic mash-up, uniform access to metadata, data, information and knowledge. Some metadata can be generated automatically (sometimes on the fly) from the tool databases according to common vocabularies like Dublin Core, SKOS, SIOC, FOAF, etc. Most of these vocabularies are lightweight ontologies that can fit well database schemas. These vocabularies provide common semantic enabling computers to put queries on LMS and social media tools. Thus, the web can be viewed as a single global database. Users and/or computers can perform complex queries against this global database using the SPARQL language. Complex queries are queries over multiples pages / web sites / data repositories whatever the tool is. It only has to expose data on the common standard and vocabularies. Thus, future pervasive learning environments can be composed of lots of different LMS, social media tools exposing, sharing and connecting data, information and knowledge on the web.
The evolution of the Internet of Things (IoT) started decades ago as part of the first face of the digital transformation, its vision has further evolved due to a convergence of multiple technologies, ranging from wireless communication to the Internet and from embedded systems to micro-electromechanical systems. As a consequence thereof, IoT’ platforms are being heavily developed, smart factories are being planned to revolution- ize the industry organization and both security and trust requirements are becoming more and more critical. The integration of such technologies within the manufactur- ing environment and processes in combination with other technologies such as Cloud Computing, Cyber Physical Systems, Information and Communication Technologies as well as Enterprise Architecture has introduced the fourth industrial revolution referred to also as Industry 4.0. In this future world machines will talk to machines (M2M) to or- ganize the production and coordinate their actions function of the information collected by different sensors and exchanged with other entities. However opening connectivity to the external world raises several questions about data security that was not an issue when devices were controlled locally and just few of them were connected to some other remote systems. That’s why ensuring a secure communication between heterogeneous and reliable devices is essential to protect exchanged information from being stolen or tampered by malicious cyber attackers that may harm the production processes and put the different devices out of order. Without appropriate security solutions, these systems will never be deployed globally due to all kinds of security concerns. That’s why ensuring a secure and trusted communication between heterogeneous devices and within dynamic and decentralized environments is essential to achieve users acceptance of such solutions. However, building a secure system does not only mean protecting the data exchange but it requires also building a system where the source of data and the data itself is being trusted by all participating devices and stakeholders.
40 4. Data Partitioning for Fast Mining of Frequent Itemset
correlations of features. Their discovery is known as Frequent itemset mining (FIM for short), and presents an essential and fundamental role in many domains. In business and e-commerce, for instance, FIM techniques can be applied to recommend new items, such as books and different other products. In science and engineering, FIM can be used to analyze such different scientific parameters (e.g., based on their regularities). Finally, FIM methods can help to perform other data mining tasks such as text mining , for instance, and, as it will be better illustrated by our experiments in Section 4.3, FIM can be used to figure out frequent co-occurrences of words in a very large-scale text database. However, the manipulation and processing of large-scale databases have opened up new challenges in data mining . First, the data is no longer located in one computer, instead, it is distributed over several machines. Thus, a parallel and efficient design of FIM algorithms must be taken into account. Second, parallel frequent itemset mining (PFIM for short) algorithms should scale with very large data and therefore very low MinSup threshold. Fortunately, with the availability of powerful programming models, such as MapReduce  or Spark , the parallelism of most FIM algorithms can be elegantly achieved. They have gained increasing popularity, as shown by the tremendous success of Hadoop , an open-source implementation. Despite the robust parallelism setting that these solutions offer, PFIM algorithms remain holding major crucial challenges. With very low MinSup, and very large data, as will be illustrated by our experiments, most of standard PFIM algorithms do not scale. Hence, the problem of mining large-scale databases does not only depend on the parallelism design of FIM algorithms. In fact, PFIM algorithms have brought the same regular issues and challenges of their sequential implementations. For instance, given best FIM algorithm X and its parallel version X ′ . Consider a very low
In practice, many games rely on a simple strategy, where players send updates at a regular rate to other players. The main flaw of this technique is a poor scalability in terms of bandwidth, as the number of messages increases quadratically with the number of players. Scalability is a concern for DVEs: some games are intended to be played by a large number of participants at the same time (e.g. MMORPGs such as World Of Warcraft). In addition, many online games are based on a client-server architecture. This has many disadvantages, as maintaining a server is often expensive, and exposes a single point of failure . This leads to the incentive to study peer- to-peer solutions, where players share the role of the server among themselves, but in this context, bandwidth becomes crucial, as the network capacities of peers are usually lower than those of powerful servers. This article focuses on reducing bandwidth usage by limiting the number of exchanged messages. Several versatile techniques have been proposed to achieve this goal.
trol and management schemes for optical networks, but these ideas are still far from practical implementation. Zervas and Simeonidou have briefly provided a prototypical distributed cognitive architecture called COGNITION [53, 52]. Cognition is expected to be implemented from the bottom physical layer to the top application layer of the network architecture across one or multiple domains. However, more implementation details and verifications are needed, especially on the complicated network manage- ment and control plane. Wei et al. have proposed a cognitive optical substrate with a mesh topology only in the core networks. This substrate aims to provide high- speed, bandwidth-on-demand and rapidly-adaptive wavelength services in a client- service-aware approach . However, the design of the substrate concerns only the physical layer and has no coordination with higher-layer functionalities. The project Cognitive Heterogeneous Reconfigurable Optical Networks (CHRON) has been pro- posed by several groups of researchers in Europe  to improve the dynamicity of the optical network control planes. The project has investigated in the intelligent monitoring techniques, the cross-layer cognitive control architecture design, and the multi-objective optimization of the performance in term of cost and energy efficiency . Monroy et al. have designed a network control plane architecture in CHRON [36, 19]. The control plane of CHRON includes a cognitive decision system and a network monitoring system. Later, de Miguel et al. have successively elaborated the centralized cognitive framework and built a testbed [20, 43, 42, 6]. However, their design has not comprehensively solved the efficient network monitoring problem or the fast reconfiguration problem. Moreover, the four-node network topology in their testbed is too simple to validate the performance of the design in real-life networks, which usually have hundreds of nodes. In particular, there is no work that shows that reconfiguration can be done based on observed traffic changes. In summary, a comprehensive cognitive network management and control scheme for current optical networks is needed.
2. Agent-Oriented Modeling and the Chemical Reaction Metaphor
The agent concept provides a focal point for accountability and responsibility for coping with the complexity of software systems both during design and execution . It is deemed that software engineering challenges in developing large scale distributed learning environment can be overcome by an agent-based approach . In this approach, a distributed learning system can be modeled as a set of autonomous, cooperating agents that communicate intelligently with one another. As an example, Collaborative Agent System Architecture (CASA) [4, 5] is an open, flexible model designed to meet the requirements from the resource-oriented nature of distributed learning systems. In CASA, agents are software entities that pursue their objectives while taking into account the resources and skills available to them.
Lack of scalability is a key issue for virtual-environment technol- ogy, and more generally for any large-scale online experience be- cause it prevents the emergence of a truly massive virtual-world infrastructure (Metaverse). The Solipsis project tackles this issue through the use of peer-to-peer technology, and makes it possible to build and manage a world-scale Metaverse in a truly distributed manner. Following a peer-to-peer scheme, entities collaborate to build up a common set of virtual worlds. In this paper, we present a first draft of the Solipsis architecture as well as the communi- cation protocol used to share data between peers. The protocol is based on Raynet, an n-dimensional Voronoi-based overlay network. Its data-dissemination policy takes advantage of the view-depedent representation of 3D contents. Moreover, the protocol effectively distributes the execution of computationally intensive tasks that are usually executed on the server-side, such as collision detection and physics computation. Finally, we also present our web component, a 3D navigator that can easily run on terminals with scarce re- sources, and that provides solutions for smooth transitions between 3D Web and Web 2.0.
Obviously, the higher the workload is built in this hierarchy, the more realistic it is. But while the high level approach may be of interest in a more comparison-oriented performance evaluation domain (as with benchmark suites, for instance), its use in a more behavior-ori- ented modeling domain leads to a few drawbacks. First, there is a high risk for workload situations built that way to cover only a subset of all the possible states of system, depen- ding on the initial pool of test applications or algorithms. In order to minimize this risk, one will have to oversize this pool, which on the other hand will also lengthen the measurement procedures and globally the whole modeling. Let us notice also that the behavior of such applications may vary from one architecture to another, depending on implementation de- tails. Consequently, it is clear that a behavior modeling based on such an analysis method is heavy, complex and hardly portable.
HAL Id: inria-00070004
Submitted on 19 May 2006
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
Mining maximally informative k-itemsets (miki ) has shown many challenges. Yet, the utility of informative patterns can be more interesting in classification field. Indeed, clas- sification is one of the building bricks in data mining and information retrieval. The problem has been widely studied in centralized environments (CE ). However, in massively distributedenvironments, parallel classification algorithms have not gained much in terms of accuracy. In this chapter, we address the problem of parallel classification in highly distributedenvironments. We propose Ensemble of Ensembles of Classifier (EEC), a par- allel, scalable and highly accurate classifier algorithm. EEC renders a classification task simple, yet very efficient. Its working process is made up of two simple and compact jobs. Calling to more than one classifier, EEC cleverly exploits the parallelism setting not only to reduce the execution time but also to significantly improve the classification accuracy by performing Two-Level Decision Making (TLDM) steps. We show that the EEC classi- fication accuracy has been improved by using informative patterns and the classification error can be bounded to a small value. EEC has been extensively evaluated using vari- ous real-world, large datasets. Our experimental results suggest that EEC is significantly more efficient.
MM’08, October 27–November 1, 2008, Vancouver, Canada. Copyright 2008 ACM 1-59593-447-2/07/0010 ...$5.00.
The multimedia research community have made much progress on audio and video transmissions, enabling high quality audio com- munications and video streaming within the NVE. The quality of 3D objects in NVEs, however, is still primitive and not realistic in general. Simplified models or image-based representations are commonly used in NVE to reduce both computational and band- width requirements. While Moore’s Law and advances in GPU technology have made concerns on computational requirements less relevant, network bandwidth still remains a bottleneck. For instance, current generation of GPU is capable of rendering the Stanford’s Thai Statue model with 10 millions triangles but the model, with a size of 122MB after compression, needs 1.6 min- utes to download even on a fast 10 Mbps link. The latency induced by downloading completely such an object during a client naviga- tion is unacceptable for interactive use. Thus, to enable realistic, high resolution 3D object in NVE, it is not feasible to render a 3D object only after it is completely received.
Job 2 Overlapping Partitions: The format of the MapReduce output is set to "Mul-
tiFileOutput" in the driver class. In this case, the keys will denote the name of each overlapping data partition output (we override the "generateFileNameForKey- Value" function in MapReduce to return a string as a key). In the map function, first, we store (once) the previous MapReduce job (Centroids) in a (key, value) data structure (e.g.MultiHashMap, etc.). The key in the used data structure is the split name, and the value is a list of items. Then, each mapper takes a transaction (line of text) from the database D, and for each key in the used data structure, if there is an intersection between the values(list of items) and the transaction being processed, then the mapper emits the key as the split name (in the used data structure) and value as the transaction of D. The reducer simply aggregates over the keys (split names) and writes each transaction of D to an overlapping data partition file. Example 20. Figure 4.1 shows a transaction database D with 5 transactions. In this example, we have two non-overlapping data partitions at step (1) and thus two centroids at step (2). The centroids are filtered in order to keep only the items having the maxi- mum number of occurrences (3). IBDP intercepts each one of these two centroids with all transactions in D. This results in two overlapping data partitions in (4) where the inter- sections only are kept in (5). Finally, the maximal frequent itemsets are extracted in (6). Redundancy is used for the counting process of different itemsets. For instance, transac- tion ef g is duplicated in both partitions in (5) where the upper version participates to the frequency counting of a and the lower version participates to the frequency counting of f g.
Membership management. Every time a swarm.create() method is executed, the BVM stores the
identifier of the created swarm into a dedicated hash table, along with a flag encoding whether the robot is a member of the swarm (1) or not (0). Upon joining a swarm, the BVM sets the flag corresponding to the swarm to 1 and queues a message <SWARM JOIN, robot id, swarm id>. Analogously, when a robot leaves a swarm, the BVM sets the corresponding flag to 0 and queues a message <SWARM LEAVE, robot id, swarm id>. Because leaving and joining swarms is not a particularly frequent operation, and motion con- stantly changes the neighborhood of a robot, it is likely for a robot to encounter a neighbor for which no information is available. To maintain everybody’s information up-to-date, the BVM periodically queues a message <SWARM LIST, robot id, swarm id list> containing the complete list of swarms to which the robot belongs. The frequency of this message is chosen by the developer when configuring the BVM installed on a robot.
4.1 Problem description
Recent developments in the automated vehicle field are facing highways and urban environments. Highways, being the safest and less challenging environments, have already been tackled by dif- ferent OEMs like Tesla, Mercedes, among others (see [Karush et al., 2016]). On the other hand, urban environments represent the most challenging scenario, because of its complexity and the number of road actors that interact with the vehicle. Different European projects have recently fo- cused in urban driving like CityMobil, CATS, V-Charge, CityMobil2, deploying automated vehicle demonstrations across Europe, as well as individual actors such as Google, Uber and NuTonomy (these last two deploying automated taxi systems in Pittsburgh and Singapore respectively—but still keeping human drivers behind the wheel for safety reasons—). Moreover, reports from the E.U. [ERTRAC, 2015] and data analysis in the U.S. [Anderson et al., 2014] show that continu- ous research in automated vehicle technology (ranging from ADAS up to fully automated) can greatly reduce emissions, travel times and risks, while at the same time increase comfort, enable social acceptance of the technology and push forward traffic optimization. However, there is still a long way before full automation capabilities avoid all road accidents as shown in recent fatal crash reports from the E.U. [European Commision, 2015] and the U.S. [NHTSA, 2016], reporting difficulties in decreasing the number of fatal accidents in European and U.S. roads respectively.