Network Coding Based Reliable Data Delivery in Large-scale Dynamic Networks

Abstract

최근 웨어러블 디바이스(wearable device), 커넥티드 카(connected car), 모바일 메디컬 기기(mobile medical devices) 등과 같이 다양한 무선 통신 기기들이 생겨나면서 우리는 언제, 어디에서, 누구에게나 연결될 수 있는 초연결(hyperconnected) 사회에 살고 있다. 과거에는 저 사양 단순 기능만 가지고 있던 여러 단말들이 계산 능력과 무선통신 기능을 가지면서 정보를 생성, 가공, 공유할 수 있게 되는 사물들(things)로 등장함에 따라 이제는 연결의 대상이 더 이상 ‘인간’으로만 한정되는 것이 아니라 주변에 존재하는 모든 ‘사물’들로 그 범위가 확장 되었다. 이러한 기술적 변화는 인간의 삶을 더욱 풍요롭게 해 주었지만, 이를 지원하기 위한 네트워크 관리 및 구축 측면에서는 해결해야 할 수많은 새로운 기술적 문제점들을 야기하였다. 먼저, 네트워크의 크기가 커짐에 따라 최적의 네트워크 토폴로지(topology)를 찾는데 필요한 계산의 복잡도가 증가하여 더 이상 기존의 중앙제어(centralized control)기반의 네트워크 디자인은 현실적인 해결책이 되지못한다. 또한, 네트워크를 구성하는 무선통신 기반의 모바일 디바이스의 수가 증가함에 따라 네트워크 환경이 더욱 다이나믹하고 불안정하게 되었다. 마지막으로, 많은 수의 디바이스들이 많은 양의 정보를 생성함에 따라 기존의 네트워크는 전송량 한계에 다다르게 되었다. 본 학위 논문에서는 이와 같은 문제점들을 해결하기 위해 대규모 다이나믹 네트워크에서의 네트워크 코딩 기반 고신뢰 데이터 전송 기법을 연구 하였다. 네트워크의 전송량을 최대화하기 위하여 네트워크 코딩 기반의 정보 전송을 제안하였고, 대규모 네트워크 토폴로지를 분산적으로 결정하고 안정적인 네트워크를 구축하기 위하여 게임 이론(game theory)과 강화학습(reinforcement learning)에 기반을 둔 해법을 제안하였다. 또한 패킷 손실이 발생하는 다이나믹 네트워크 환경에서도 신뢰성 높은 정보전송을 위하여 전송 정보의 상관성(inter-dependency)을 활용한 근사 복호법 (approximate decoding)과 압축센싱기법(compressed sensing)을 활용한 압축네트워크코딩 기법(compressed network coding)을 제안하였다. 이를 통해 네트워크 코딩의 한계점으로 알려진 All-or-nothing 문제를 해결 할 수 있게 되었다. 마지막으로, 앞에서 제안한 네트워크 코딩 기반 정보전송 기술을 실용화 된 IPTV 네트워크와 실시간 멀티미디어 전송 시스템에 적용 및 구현하는 것으로 본 학위 논문을 마무리 하였다.;Increasing number of connected devices, such as wearable devices, con- nected cars, unmanned aerial vehicles, and mobile medical devices have led to the hyperconnected world where anybody can be connected anytime and anywhere. Not only with other human beings but also with the things around us are interconnected, while generating, processing, utilizing and exchanging information; of which a growing number have complex computational capabil- ities and wireless connectivity. Such advances on a networked system create an opportunity to design an unprecedented array of new applications and ser- vices. However, there are several challenges in across the entire spectrum of network architectures. First of all, as the network size becomes larger, designing the optimal topol- ogy is much more complicated such that conventional centralized network de- sign solutions cannot be practically considered anymore in a large-scale net- work. Moreover, pervasive computing environment draws increased network dynamics where network nodes with high mobility cause frequent changes in members associated with the network and unstable channel conditions with high link failure rates. Furthermore, significantly large amount of information in pervasive computing environment causes limited throughput in network. These challenges in recent network demand new approaches for reliable and robust data delivery. This dissertation addresses the abovementioned challenges by developing network coding based framework for large-scale dynamic networks. We deploy network coding for efficient information delivery as it can utilize maximum throughput of networks. Then, we first study distributed network topology design for a large-scale network in conjunction with network coding based on two approaches: game-theoretic approach and reinforcement learning ap- proach. By considering the nodes in the network as agents that can take ac- tions for making connection to other nodes, we show that a stable network can be formed based on a set of actions that is in the Nash equilibrium at each agent. This is referred to as the network formation game that consists of a large number of players and multicast fl ws, and thus it requires high computational complexity to fi the solution to the game. We show that the network formation game can be decomposed into independent link for- mation games played by only two players with a unicast fl w, which requires significantly low complexity. Hence, a stable topology of a large-scale network can be designed with low computational complexity in a distributed way. We next focus on wireless networks where the topologies can be determined by the decisions on wireless transmission ranges of each node. In order to maxi- mize throughput gain by deploying inter-session network coding where linear network coding cannot always achieve capacity while sometimes no coding can achieve, we include the option for taking network coding operations in an action set of a node. Lastly, we further improve the proposed strategy by deploying the Markov decision process such that the network topology can be adaptively evolved against network dynamics. Each node fi its optimal pol- icy which returns the optimal changes in transmission range given the number of nodes that successfully receive data. This is referred to as effective nodes. The proposed strategy can maximize the expected long-term utility, which is achieved by considering both current network conditions and future network dynamics. We show the resulting network eventually converges to stationary networks maintaining a determined number of effective nodes in transmission range. We also propose how to initialize the network such that the convergence of the proposed strategy can be expedited. Once network topology is designed or given in an application, our focus should be on robust and reliable data delivery with network coding. For delay- sensitive applications, in particular, network coding may cause all-or-nothing problem where the decoder cannot reconstruct any encoded data unless they receive at least as many coded packets as the original number of packets within time constraints. This leads to failure in successful data delivery. In order to overcome the limitation and provide uninterrupted data delivery, we propose approximate decoding strategies that enable decoders to perform re- construction process with insuffi t set of data. We show that the perfor- mance of the approximate decoding strategies can be improved by exploiting inter-dependency of sources. In particular, the diff between source vec- tors is characterized by a unimodal distribution, we propose a mode-based algorithm for approximate decoding, where the mode of the source data distri- bution is used to reconstruct source data. We further improve the mode-based approximate decoding algorithm by using additional short information that is referred to as position similarity information. As an alternative solution for all- or-nothing problem, we also propose compressed network coding which jointly considers network coding and compressed sensing. With the help of compressed sensing, the decoder is able to approximately recover the source data based on l1-norm minimization. The last part of dissertation includes implementation of network coding in a commercial product. We fi propose an architecture for IPTV networks where network coding is deployed in the backbone networks. As the number of subscribers to IPTV services increases, the size of backbone network is correspondingly grows and it is challenging to manage network faults in a centralized manner. Moreover, the size of video contents in IPTV services rapidly grows such as ultra high defi (UHD), 4K UHD, and 8K UHD, the limited network capacity becomes a critical problem of current networks. In order to overcome such challenges of IPTV services in the current networks, we propose to deploy network coding technique, which can increase network capacity while improving robustness against network faults. We next develop a real-time multimedia streaming system based on network coding, where a large number of clients are considered in crowded area (e.g., live video streaming at a sport stadium). Conventional acknowledgement based approaches cannot support reliable transmission in such scenarios in practice because the server may not be able to control all feedback channels from a large amount of clients. Instead, the proposed approach can provide a novel error concealment strategy by deploying systematic network coding. In our system, the broadcast server transmits “healing packet” (i.e., network coded packets) as well as the original source packets. Hence, the receiver can still reconstruct the lost packets by using healing packets, even if the receiver misses some of original packets. In summary, this dissertation covers entire spectrum of network architec- tures: distributed network formation for a network coding enabled large scale dynamic network, reliable and robust data delivery based on network coding, and practical development of network coding in commercial products. The pro- posed approaches enable us to establish a better hyperconnected world with a large number of connected entities in dynamic environment.