Search

Collision avoidance is one of the crucial functions of autonomous driving systems (ADSs) to guarantee the safety of autonomous vehicles (AVs). It requires extensive testing before an AV is deployed to public roads. Most of the current ADS testing methods generate test cases either from real traffic data or manually designed for some specific scenarios. There is little work on systematic methods to generate test cases from a complete map where an AV operates. Systematic testing on such a map is challenging due to the enormous scenarios. In this paper, we propose a collision-avoidance testing method for ADSs running on a map, which aims to reduce the scenario space while maintaining scenario diversity. The method consists of test case classification and test case generation. First, we build the topology structure of a map, based on which we classify possible scenarios into different classes. Second, we divide test cases into different classes using the topology-based scenario classification and fuzzy number-based motion evaluation. Third, we implement a bisection method to generate test cases that can efficiently expose ADSs' failures. We evaluate our method on one of the state-of-the-art ADSs, Baidu Apollo. The experiment results show that our method discovers Apollo's issues effectively while reducing the number of generated test cases by 77.36%, compared with the random method.

Autonomous vehicles (AVs) play an important role in transforming our transportation systems and relieving traffic congestion. To guarantee their safety, AVs must be sufficiently tested before they are deployed to public roads. Existing testing often focuses on AVs’ collision avoidance on a given route. There is little work on the systematic testing for AVs’ route planning and tracking on a map. In this paper, we propose CROUTE, a novel testing method based on a new AV testing criterion called route coverage. First, the map is modeled as a labeled Petri net, where roads, junctions, and traffic signs are modeled as places, transitions, and labels, respectively. Second, based on the Petri net, we define junctions’ topology features and route features for junction classification. The topology feature describes the topology of roads forming the junction, and the route feature identifies the actions that a vehicle can take to follow a route. They can characterize route types on a map. Hence, route coverage measures how many route types are covered. We then propose a systematic method that aims to cover all route types for a well-designed AV system with a small number of test cases. We implement and evaluate CROUTE on Baidu Apollo running with the LGSVL simulator. We carry out testing on the map from a section of San Francisco and find six different types of issues in Apollo. The experiment results show the validity of route coverage and the efficiency of CROUTE.

The traditional lead zirconate titanate ceramic (PZT) bimorph deformation is small, it is difficult to be used as MAV morphing wing actuator. Thus, a cantilever axial compression macro-fiber composite (MFC) bimorph is proposed in this paper. Firstly, the theoretical model and linear finite element method (FEM) model of MFC bimorph are established based on the homogenization parameters of MFC. The inactive part of MFC and nonlinear tensile stiffness of MFC are unconsidered in these models. Then, the nonlinear tensile stiffness of MFC is introduced into the FEM model of whole bimorph to obtain the deformation characteristics and design space. The result shows that the free end displacement of 12.7 mm can be achieved in the case of the driving voltage of 1440/-480V with the axial compressed force of 8 N, which is three times larger than the same bimorph without axial compressed force. The corresponding design space has been expanded three times.

This paper proposes to train a group of badminton serving machines by a simple deep neural network (DNN) consisting of 8 dense layers. Our ultimate goal is to have the proposed system reproduce a badminton rally captured from social media such as YouTube. All the parameters of these serving machines need to be predicted from a trajectory image in a race video through a deep learning regression model. We can successfully estimate 10 trajectory parameters with an average RMSE loss of 0.087.

People have relatively higher requirements for image storage in some specific fields, such as high-resolution cultural relic data image, medical image, infrared remote sensing image, high-precision astronomical observation image. There cannot be any pixel loss in the storage process, so the image can only be compressed by lossless compression. In this paper, a lossless image compression algorithm based on the neural network of long short-term memory (LSTM) is proposed: a LSTM model predictor based on attention mechanism is built by utilizing the memory characteristic of cyclic neural network. The previous pixel value of the image was taken as the input of the model, then the predicted pixel was obtained through the cyclic neural network, and finally the calculated difference between these values was encoded by the mixed run-length encoding and Golomb-Rice encoding. Compared with the traditional predictive lossless image compression algorithm, this algorithm proposed here comprehensively considers the correlation between more pixels and encoded pixels. The experimental results show that compared with the lossless image compression algorithms JPEG-LS and CALIC, the proposed algorithm improves the compression rate by 25% and 12% respectively.

Out of vocabulary (OOV), which is a word that does not exist in a predefined vocabulary. How to deal with OOV is an important research topic in the field of natural language processing. The existence of OOV directly affects the performance of many NLP systems. For example, in some common scenarios such as machine translation, sentiment analysis, and intelligent question answering, the existence of OOV can greatly affect the key performance of the system. In recent years, with the advent of the word vector algorithm word2vec based on the principle of word morphology, the word embedding path of the NLP system has improved significantly. We combine LSTM with NLM, taking the morphemes of words as the basic processing unit, while taking into account the global context information. The results obtained are better than the existing OOV processing strategies, and the performance of commonly used NLP systems is generally improved. Finally, it is experimentally proved that our model is generally better than the existing models in the problem of unregistered word processing.

This paper presents the electrical behavior of normally-off AlGaN/GaN HEMT under heavy ion irradiation based on technology computer aided design numerical simulation to better understand single-event burnout (SEB). Firstly, the effects of heavy ion impact at different locations on the device are simulated. Secondly, more complex radiation angles are introduced to find the effects of different angles on the electrical characteristics of the device. Finally, the most sensitive conditions with respect to SEB failure are obtained.

In order to extract the effective features of EEG data based on P300 potential and improve the classification accuracy of brain-computer interface (BCI) system, a novel method from spatial and frequency domain is proposed in this paper. That is, the multi-rhythm signal and Common Spatial Pattern (CSP) are combined to extract the feature of P300 potential. Meanwhile, this paper uses the elastic networks as classifier which combines with $l_{1}$ norm and $l_{2}$ norm that not only makes the coefficients sparse, but also avoids losing the inhere t structure among samples from the same class. In this paper, the public EEG data set is used to verify the proposed method and by comparing the traditional CSP method. The experimental results show that the new method provides more frequency-space-related feature information increased by 4.8% compared with the traditional CSP method. It indicates that the proposed method achieves the good performance for the P300 EEG analysis.

Massive Open Online Courses (MOOCs) trade deep and intimate tutor-tutee interactions for large-scale coverage of educational lectures. Therefore, a proper length of student-assistant discussion is particularly precious both for the questioner and the observers in MOOC communities. To investigate the concept conveying characteristics of long and short discussions, this paper collected 32 student-assistant discussion transcripts in MOOC communities. Each discussion thread was randomly chosen in one course. Besides, 32 questions-and-answers (Q&A) transcripts in different science lectures and 31 Q&A transcripts in different press briefings were collected as comparison groups. The nouns in these transcripts were extracted as concepts, and the concept conveying procedure was modelled into a directed graph. We found that if an assistant would like to prolong the discussion in MOOC communities, he or she needs to retain concepts closely connected like political spokesmen on press briefings, and apply properly deep concepts like senior scientists on science lectures.

This paper analyzes the factors affecting the fixed pricing of real-time data products mainly through customer perceived value and data score, using data visualization tools and regression assumptions. The article uses the complex network analysis software GEPHI to quantitatively analyze and draw the visual spectrum of the relationship between the price of realtime data products and the factors influencing the possibility, and conduct general regression analysis on this basis. Through a dual argument, we get the fixed price of real-time data products key influencing factors in order to provide constructive solutions for the pricing of data products.

Matrix computation is the core of many massive data-intensive analytical applications such mining social networks, recommendation systems and nature language processing. Due to the importance of matrix computation, it has been widely studied for many years. In the Big Data ear, as the scale of the matrix grows, traditional single-node matrix computation systems can hardly cope with such large data and computation. Existing distributed matrix computation solutions are still not efficient enough, or have poor fault tolerance and usability. In this paper, we propose Marlin, an efficient distributed matrix computation library which is built on top of Spark. Marlin contains several distributed matrix operation algorithms and provides high-level matrix computation primitives for users. In Marlin, we proposed three distributed matrix multiplication algorithms for different situations. Based on this, we designed an adaptive model to choose the best approach for different problems. Moreover, to improve the computation performance, instead of naively using Spark, we put forward some optimizations including taking advantage of the native linear algebra library, reducing shuffle communication and increasing parallelism. Experimental results show that Marlin is over an order of magnitude faster than R (a widely-used statistical computing system) and the existing distributed matrix operation algorithms based on MapReduce. Moreover, Marlin achieves comparable performance to the specialized MPI-based matrix multiplication algorithm SUMMA but uses a general dataflow engine and gains common dataflow features such as scalability and fault tolerance.

In a new era of Big Data, the rapid growth of the applications, such as social media and web-search, requires efficient and scalable machine learning and statistical analytical algorithms. However, there lacks easy-to-use and efficient software frameworks or systems that can support fast development of such big data analytical algorithms. To solve these problems, we propose Octopus, an easy-to-use and efficient analytical system for big data. Octopus allows data analysts conduct complex data analytics for big data with traditional programming languages and methods in an easy and efficient way. To achieve the goal of ease-to-use, we propose a matrix-based unified programming model, which is the core of many data-intensive statistical applications such as numerical analysis and data mining. Further, in order to improve the performance, the Octopus software framework adopts various distributed computing platforms, including Hadoop MapReduce, Spark and MPI. On these computing platforms, we design several parallel matrix computation algorithms, which are suitable for various scenarios. Finally, the features of Octopus are encapsulated into a library with matrix-based APIs and exposed to users as an R package. R is a widely-used statistical programming language and supports diversified data analysis tasks through extension packages. Experimental results show that Octopus achieves efficient performance and near linear scalability.

As an important cultural factor, individual religious belief how to impact his/her psychology and behavior need more attention and researches, especially in such a complicated religious context of China. Based on related literatures and Chinese conditions, an exploratory research about the relationship between Chinese religiosity and their customer loyalty was carried out. At first the new theory of consumer religiosity including its definition, content and measurement was introduced and reviewed. Then a new scale to measure Chinese religiosity is established. With statistical analysis and survey data, the significant relationship between religiosity and customer loyalty was found and shown to be positive. This contributes a lot to the academic area of religion and consumer behavior, as well as corporates' loyalty programs and customer retention.

Output variable falls in a specified interval with a certain probability control problem for the output stochastic distribution control (SDC) systems, where those systems associate with the model parameters uncertainty and random disturbance, is considered in this paper. This problem can be transferred into a robust tracking control problem with a distribution tracking errors upper bound constrained by the output PDFs being approximated B-spline neural network. A controller design method, linear matrix inequality (LMI) method, is used to this output SDC system. Finally, Simulations are included to show the feasibility and effectiveness of the proposal control algorithm, and encouraging results have been obtained.

A method for speaker normalization in deep neural network (DNN) based discriminative feature estimation for automatic speech recognition (ASR) is presented. This method is applied in the context of a DNN configured for auto-encoder based low dimensional bottleneck (AE-BN) feature extraction where the derived features are used as input to a continuous Gaussian density hidden Markov model (HMM/GMM) based ASR decoder. While AE-BN features are known to provide significant reduction in ASR word error rate (WER) with respect to more conventional spectral magnitude based features, there is no general agreement on how these networks can reduce the impact of speaker variability by incorporating prior knowledge of the speaker. An approach is presented in this paper where spectrum based DNN inputs are augmented with speaker inputs that are derived from separate regression based speaker transformations. It is shown the proposed method could reduce the WER by 3% relative to the best speaker adapted AE-BN CDHMM system.

To investigate the significance of metastasis and invasion of MMP-9 and BRMS1 in breast cancer stem cells. The tissues from 65 breast cancer patients were divided into two groups, with and without lymph node metastasis, the CD44+CD24and Non-CD44+CD24- cells were separated from MCF-7 by immunomagnetic beads. ALDH1 and mRNA levels were measured by RT-PCR in different groups. The ability of metastasis and invasion in different group cells were detected by Transwell. MMP9-BRMS1 and ALDH1 were detected in different groups by immunnohistochemical staining. The level of ALDH1 mRNA in CD44+CD24- group was higher than other groups. More cells in CD44+CD24- group through the membranes than other two groups. Western blot showed that MMP9 and ALDH1 levels in CD44+CD24- group were higher than Non-CD44+CD24- cells but BRMS1 was lower (P

Based on the theory of social approval, this paper divides the home products consciousness into recessive, dominant and the direct situation. With college students as samples, this paper investigate their three situations of the home products consciousness. Through the statistical analysis and tests, The result shows the relationship between the patriotic emotion of direct situation and domestics preference of dominant situation is not evident; the relationship between the domestics preference of recessive situation and domestics preference of dominant situation is not evident; the relationship between the patriotic emotion of direct situation and domestics preference of recessive situation is negative correlation. Finally, this paper discusses the results of research, and points out the theoretical contributions, limitations of the research and the direction of the research in the future.

In view of inefficient charging and discharging management and monitoring in the intrinsic safety mode DC UPS power underground in mining at present, this paper puts forward a three-segments mode charging control strategy for the Lithium-ion battery of UPS using the microcomputer and the method for no discharging while low battery voltage. At the same time, this UPS can transmit the working parameters of its own and its battery. It also provides detailed intrinsic safety circuit with two stages voltage limiting and current limiting by using LM723, and software flow chart of microcontroller. In practice it proves that the service life of the UPS's battery be prolonged and the control center also can in time understand the state of this power.

This paper investigates a new approach for detecting phoneme level mispronunciations from utterances obtained from impaired children with neuromuscular disorders. This new pronunciation verification (PV) approach is obtained from the subspace based Gaussian mixture model (SGMM) based pronunciation model, where a set of state level projection vectors is applied for representing phonetic variability. SGMM models are trained from disabled speakers' utterances and PV scores are computed directly from distances between disabled and reference speaker projection vectors. An experimental study was performed to evaluate the performance of the SGMM based approach with respect to an approach based on the lattice posterior probabilities. A reduction in equal error rate (EER) of approximately 15% was obtained when the SGMM based scores were combined with lattice posterior probabilities.

This paper concentrates on the empirical research, which focus on scientific project performance evaluation of Zhejiang colleges'. Firstly, based on the assessment model and empirical data, it gets weights and scores of indexes. Secondly it compares every index from different type of scientific project. Thirdly, this paper analyses causes result of these outcomes. At last, it draws inspirations from analyses and brings some proposes.

This paper investigates the impact of subspace based techniques for acoustic modeling in automatic speech recognition (ASR). There are many well known approaches to subspace based speaker adaptation which represent sources of variability as a projection within a low dimensional subspace. A new approach to acoustic modeling in ASR, referred to as the subspace based Gaussian mixture model (SGMM), represents phonetic variability as a set of projections applied at the state level in a hidden Markov model (HMM) based acoustic model. The impact of the SGMM in modeling these intrinsic sources of variability is evaluated for a continuous speech recognition (CSR) task. The SGMM is shown to provide an 18% reduction in word error rate (WER) for speaker independent (SI) ASR relative to the continuous density HMM(CDHMM) in the resource management CSR domain. The SI performance obtained from SGMM also represents a 5% reduction in WER relative to subspace based speaker adaption in an unsupervised speaker adaptation scenario.

High temperature air combustion (HTAC)technology is a high efficient, low pollution of combustion technology being widely studied. The mathematical model for HTAC is built and the effect of recirculated flue gas position on combustion and NOx emission is investigated by using numerical simulation. The results show that recirculated flue gas can effectively reduce the oxygen concentration in combustion air. With the change of recirculated flue gas position, the furnace temperature and NOx concentration are changed, and the both appear peak values when the position is at the middle section of furnace. The beginning section of the furnace is an ideal position of recirculated flue gas to be chosen for low NOx emission, however, the end section is good for temperature uniformity.

In the recent years the complex networks theory as a powerful tool has been used in the study of such hot topics as internet, systems biology and optimization of traffic system. This article focuses on a network system consisting of famous repressilators coupled by quorum sensing, where intracellular and extracellular signals are integrated in an ANDN logic manner. A local synchronization criterion is established, where the cis-regulatory input function of ANDN response plays a key role. Moreover, the local result is extended to the global case through the specific structure of the network.

There is an increasing demand for network devices to perform deep packet inspection (DPI) to enhance network security. In DPI the packet payload is compared against a set of predefined patterns which can be specified using regular expressions (regexes). It is well-known that mapping regexes to deterministic finite automata (DFA) will suffer from the state explosion problem. Through observation, we attribute DFA explosion to the necessity of remembering matching history. In this paper, we investigate how to record the matching history efficiently and propose an extended DFA approach for regex matching called fcq-FA, which can make a memory size reduction of about 1000 times with a fully automated approach. In fcq-FA, we use pipeline queues and counters to help recording the matching history. Hence, state explosion caused by Kleene closure and repetitions can be definitely avoided. Further, it achieves a fully automated signature compilation with polynomial running time and space.

In Network Intrusion Detection System, De-terministic Finite Automaton (DFA) is widely used to compare packet content at a constant speed against a set of patterns specified in regular expressions (regex patterns). However, combining many regex patterns into a single DFA causes a serious state explosion. Partitioning the pat-tern set into several subsets, each of which produces a small DFA, is a practical way to deflate the state explosion. In this paper, we propose a regex pattern grouping scheme based on a new DFA model called Pattern-Based DFA (P-DFA) which supports efficient pattern-based op-erations, such as insertion, deletion, and etc. By using these basic operations, one can easily measure the state explo-sion when combining a set of regex patterns into a single DFA. Based on the privilege, we develop regex grouping algorithms for mitigating the state explosion in parallel and sequential matching environments, respectively. The evaluation shows that under the same constraints, our ap-proach requires only half the number of groups compared with the most well-known algorithms.

This paper theoretically researches the application that ultrasonic standing waves are used to generate force on particles within fluid and the trajectory of the particles is given with numerical solutions. Based on this theory a device has been fabricated in pyrex to manipulate particles which are concentrated in the cavity and an experiment system is established. The experiment was accomplished on the device with 10µm diameter polystyrene spheres. The result was accordant with the theory. When the piezoelectric transducer on both sides of the device turned on, the ultrasonic standing waves were formed in the cavity then polystyrene spheres moved to the wave node in several seconds. All the enrichment of polystyrene spheres was observed with the camera of the experiment system.

Web services have become the de facto standard as component technologies for quickly composing a business process that satisfies the business goal of an organization. Nowadays, Web services have found its way into describing the functions of automatic tasks as well as human tasks. An important part in the specification of a business process, especially for human tasks, is the access control that constrains who can perform its tasks. This paper considers both types of tasks involved in a Web services-based process and the corresponding access control problem. We focus on two types of access control constraints, namely separation of duties (SoD) and binding of duties (BoD). Both role-level and participant-level of SoDs and BoDs are investigated in this paper. We propose an approach to dynamically choosing the performer for each task so as to satisfy all access control constraints. The proposed performer selection approach is evaluated based on a purchase order scenario and is shown to have the highest chance of satisfying all predefined access control constraints when compared to other methods.

New applications such as real-time deep packet inspection require high-speed regular expression (regex) matcher, and the number of regexes in pattern store is increasing to several thousands, which requires a memory efficient solution. In this paper, a kind of hardware based compact DFA structure for multiple regexes matching called CPDFA is presented. According to statistics of regexes in Snort and L7-filter rules, transitions from each state to its next states are not evenly distributed. The summation of transitions from each state to its top three most popular next states takes about 90% of all the transitions. Therefore, CPDFA employs an indirect index table to represent transitions to top three most popular next states more efficiently. The remaining transitions which take about 10% of all the transitions are stored in direct transition table or K parallel SRAMs according to the number of remaining transitions from the same state is more than K or not. Simulation shows that CPDFA structure can save about 90% of memory storage comparing with the original DFA structure. By using pipelined architecture in FPGA, CPDFA can advance one character in one memory access cycle.

Despite the popularity of streaming applications over Internet, little is known about its practical properties which could be potentially exploited to guide the system design of next generation peer-to-peer (P2P) networks. In this paper, we hence investigate the problem of understanding and supporting the two types of dominant streaming services - live broadcasting and on-demand streaming - with high service availability and system scalability. We make a measurement study to traditional but popular client/server (C/S) systems to reveal many fundamental and interesting observations with the benefit of totally more than 30,000,000 real workload traces. The anatomy particularly focuses on request and popularity issues which play a vital role to the system performance and also characterize evolutional user behaviors in the service community. The inherent request characteristics in live broadcasting service and drawbacks of classical cache-and-relay (CR) scheme in on-demand streaming further motivate us to discuss essential challenges and open questions in the context of supporting the two services over P2P networks.

In this paper we discuss a kind of optimal control problems with restricted control variables, and the state equations of which are depicted by differential-algebraic equations(DAEs). Through analyzing characters of the optimal control model with DAE constraints, the model can be transformed into a functional extreme value problem. Then by means of the variational principle, we present necessary conditions of optimality for the functional extreme value problems. In addition, a minimum condition of optimality is given based on the necessary conditions. Those theoretical results also provide an analytical way for solving optimal control problems(OCPs). Finally, a well-known chemical engineering model is shown to explain our results.

Using principal component analysis (PCA) and improved fuzzy neural network by PSO to evaluate the success degree in electric power engineering is this paper's innovative points. First we construct the algorithm model which based on PCA and BP neural network improved by PSO. Secondly using PCA to predigest the given index system and then using the relative membership degree processing the date, which as the input sample of neural network. Thirdly, use the improved BP neural network by PSO to evaluate the success degree of electric power engineering. The result denotes that it is more accuracy and speedily than BP neural network algorithm. Lastly, we give a real engineering, and get a satisfaction result.

Based on the analysis and comparisons of complexity between stochastic segment model (SSM) and hidden Markov model (HMM) in this paper, we presented a fast and robust SSM, which yields a 94.75% speaker-independent performance on Mandarin digit string recognition. This result is better than HMM based system at the same level of computational complexity and just only a little slower than HMM in the running time. We also studied a region based discriminative method, which achieves 18.0% error rate reduction for substitution error and 95.08% accuracy for Mandarin digit string recognition.

This paper presents a new approach to feature analysis in automatic speech recognition (ASR) based on locality preserving projections (LPP). LPP is a manifold based dimensionality reduction algorithm which can be trained and applied as a linear projection to ASR features. Conventional manifold based dimensionality reduction algorithms are generally restricted to batch mode implementation and it is difficult in practice to apply them to unseen data. It is argued that LPP can model feature vectors that are assumed to lie on a nonlinear embedding subspace by preserving local relations among input features, so it has a potential advantage over conventional linear dimensionality reduction algorithms like principal components analysis (PCA) and linear discriminant analysis (LDA). Experimental results obtained on the Resource Management (RM) data set showed that when LPP based dimensionality reduction was applied in the context of mel frequency cepstrum coefficient (MFCC) based feature analysis, a significant reduction of word error rate (WER) was obtained with respect to standard MFCC features.

During recent years, the Internet has witnessed a rapid growth in deployment of peer to peer (P2P) based live media streaming systems. In this paper, we are motivated to investigate the problem of improving the streaming quality of service (QoS) of those systems by two universal recognitions: one is that the understanding to practical service experiences would benefit the performance enhancement, while the other is "the more, the better" design philosophy in P2P networks. As an interesting approach, we first retrieve the practical online duration information of end users from service traces in our GridMedia system, and then intentionally extend the online duration of each peer, which would result in a more stable peer overlay. The comparative simulations with original real-world traces and modified ones demonstrate that the quality of streaming service would be better if peers stay longer in the community. We claim that, this simple but positive result essentially validates both the influence of end users' behaviors and the need of an incentive to encourage users to spend more time in P2P live streaming systems.

The Internet has witnessed a rapid growth in deployment of gossip-based protocol in many multicast applications. In a typical gossip-based protocol, each node independently exchanges data with its neighbors, acting as dual roles of receiver and sender to facilitate scalability and resilience. However, most of previous work in this literature seldom considered cheating issue of end users, which is also very important in face of the fact that the mutual cooperation inherently determines overall system performance. In this paper, we mainly investigate the dishonest behaviors in decentralized gossip-based protocol through extensive experimental study. Our original contributions come in two-fold: In the first part of cheating study, we analytically discuss two typical cheating strategies, that is, intentionally increasing subscription requests and untruthfully calculating forwarding probability, and further evaluate their negative impacts. The results indicate that more attention should be paid on defending cheating behaviors in gossip-based protocol. In the second part of anti-cheating study, we propose a simple receiver-driven measurement mechanism, which evaluates individual forwarding traffic from the perspective of receivers and thus identifies cheating nodes with high incoming/outgoing ratio. The experiments under various conditions show that it performs quite well in case of serious cheating and achieves considerable performance in other cases.

IPTV is highly advocated in recent years. Today, the growth of the residential users makes it possible to broadcast live video content through global internet to a great population of end users. In this paper, we mainly describe the design and deployment experiences of a peer-to-peer (P2P) based IPTV system called Gridmedia since P2P is universally recognized as a scalable and cost-effective means for group communication. In order to reduce the playback latency at end users as well as offer resilience to high churn rate in peer community, Gridmedia organizes end users into an unstructured overlay while proposes an efficient push-pull streaming mechanism to distribute video segments among partner nodes. To further verify the prototype, on one hand we conduct extensive experiments over PlanetLab to confirm the overlay architecture and streaming mechanism in Gridmedia can greatly reduce the average playback latency as compared to previous works; on the other hand the practical deployment of Gridmedia system over global Internet also demonstrated it could reliably support large number of users without need of network infrastructure. We also discuss, in a practical way, the issues when perfecting a prototype to a publicly released platform and system performance during live broadcasting CCTV Spring Festival Gala 2006.

Overlay network has emerged as a popular alternative to traditional client-server architecture for the distribution of media content. Intrinsic to the notion is a 'self-growing' community of end users. It is thus important to investigate practical issues in reality rather than following prescribed theoretical assumptions. This paper presents our research experiences on deployment of overlay network for live video streaming over Internet. In our implementation we adopt a gossip-based overlay construction protocol to accommodate topology changes and a composite scheduling mechanism to stream video contents. More importantly, with the service to totally more than 500,000 users and maximum 15,239 concurrent users provided by one common streaming server, we then exhibit insightful statistical results to reveal system performance and nodes properties in terms of capacity evolution, quality of service, request rate and online duration.