Professor Henry Fuchs from University of North Carolina at Chapel Hill gave the keynote speech titled "Experiences Building Tele-presence System". The keynote first explained the term tele-presence, of which the main aim of is to give users convincing immersive experience. Henry then introduced tele-presence systems that have been developed since 1997, underlining the main challenges in developing such systems. The keynote concluded by highlighting two key system challenges for immerse environment: (i) a middleware, and (ii) a set of integrated OS components, for common, lower level tele-presence functions, such as display management.

The keynote was interleaved with many questions from the audience. Kevin Almeroth (UC Santa Barbara) asked why a new term ``tele-presence'' is coined (Answer from Henry: to emphasize human "sense"). Yih-Fan Chen (Robin, AT and T Research) argued that boundary exists naturally between the virtual world and the real world and true immersion cannot be achieved, to which Henry agreed. Current solutions to virtual world are either not fully immersive or require wearing of special devices.

The role of using large displays in tele-presence systems were discussed with questions from Dick Bulterman (CWI) and Damien Le Moal (Hitachi). Henry pointed out that using large display helps users in knowing the relative position and context within the data displayed, and allows users to move around more freely. Dick also asked about the relative importance of audio and visual in tele-presence system. Henry pointed out that the importance of audio/visual information is a psychological issue and depends on the applications.

Damien asked why OS integration is important in large display system when we can just develop dedicate applications. Henry argued that system tasks (such as windows management) need to be adapted to the new display system, so only an OS supporting the large display system can give users uniform experience in both system tasks and dedicate applications.

Padmanabhan Pillai (Intel Research) and Damien asked about the role of haptic and smell. Henry answered that the sense of touch and smell are important in some cases (e.g., medical applications) but the current technology is limited. Haptic devices confine movement to a small distance, and the basis of smell (like RGB for color) are yet to be found.

Wanmin Wu from UIUC presented a 3D video collaborative system that uses an P2P overlay for dissemination of content streams. Each peer has the capability of producing and rendering multiple 3D video streams (depth as the added dimension). The main contribution of this work is an algorithm that accounts for the contributing factor of streams to the overall view when selecting peers. Experimental results demonstrated that the algorithm performs more efficiently when compared to using view proximity only.

The ensuing discussion on 3D video capturing systems revealed that the current camera technology is very limiting in capturing larger areas, as those currently being developed are being aimed at desktop usage, with ranges of 1 - 3 meters. Wei Tsang Ooi (National University of Singapore) asked about the evaluation of visual quality of 3D video, to which Wanmin replied that designing an objective visual quality metric for 3D video remains an open challenge.

Song Lin (Tsinghua University) presented analysis of characteristics of captured traffic generated by four online map applications. The findings of note are (i) the traffic follows diurnal cycles, (ii) the majority of map application traffic consists of image traffic (68.7% for Sogou), (iii) 40% of Google Maps's traffic is from mash-up sources. Due to map hotspots and the applicability of Zipf's law, the study suggests that CDN servers can improve the performance of online map application at these hotspots, especially when the hotspots would differ according to physical location of users.

Wei Cheng (National University of Singapore) asked whether there was any data supporting the conclusion that requests from different physical locations have different map hotspots. Lin replied that data supporting this is difficult to gather, but the conclusion can be inferred from the visual overlay of requests distribution on a geographical map. Wei Tsang inquired what could online map service providers learned from this work. Lin mentioned two lessons. First, providing good API is critical in encouraging other website to mash-up the map data and drive traffic to the map server. Second, CDN and other caching methods are still very relevant in such online applications.

Yuen Feng (University of Toronto) presented, on behalf of the authors, work on modeling a real world peer-assisted online storage system called FS2You, which has become popular in China with 500 GB to 1 TB of data transferred daily. The goal of the work was to derive a lower bound for the average file download time and evaluate the performance of different bandwidth allocation strategies. Using average file download time and peer satisfaction level as metrics for evaluation, it was found that FS2You's current strategy of allocating bandwidth inversely to file popularity has favorable performance, but not when bandwidth is unconstrained.

Wei Cheng asked whether the file popularity in FS2You follows the Zipf distribution, to which Feng replied that it does. Robin asked about the assumption that peers have uniform upload/download rate. Feng explained that this assumption is a simplification to the model and is an area for future work.

Hassan Shojania (University of Toronto) presented the paper on behalf of the authors. An on-demand P2P streaming systems have to handle peers with different playback positions and, consequently, a low block overlap among these peers. Hassan described InstantLeap, a mechanism to quickly discover appropriate peers for a given playback position, improving current systems with a typical O(log(#peers)) communication overhead to O(1) with high probability using a lightweight indexing protocol. The protocol groups peers by their playback locality into segments. Peers then maintain some connections to other peers from the same segment and to some peers from other segments using random exchanges of known peers.

Nazanin Magharei (University of Oregon) presented her work on clustering of peers and their impact in peer-to-peer live streaming systems using a random mesh as P2P overlay. Clusters occur, for example, due to good connectivity among peers having the same Internet Service Provider (ISP). Peers in a clusters may suffer from poor content availability in their neighborhood as the cluster's peers are mostly connected among each other and only to few outside peers. The authors provide a mechanism to detect and break-up such clusters. The proposed protocol groups peers by their distance to the source and increases the number of connections to other peers and swaps parents in case of problems. This allows reduction of the buffer size, to improve playback quality, and to reduce the distance to the source.

Michela Meo (Politecnica di Torino) presented their work addressing the problems of heterogeneous bandwidth availability in P2P live streaming systems. Like in the preceeding talks of this session, a random mesh is used as overlay. The proposed protocol better exploits high-bandwidth peers by increasing their connectivity and by positioning them closer to the source. This way, the high-bandwidth peers provide a backbone to quickly disseminate new data blocks, benefiting all the peers. The bandwidth of a peer is indirectly estimated based on the fraction of used neighbor connections and without the peer having to explicitly know or measure its bandwidth.

A short panel discussion with all the presenters was held at the end of this session. The main topic of the discussion was the locality-awareness of the proposed protocol, i.e., whether they are aware of the underlying physical topology in order to avoid long distance connections and to keep most traffic within the same ISP or close-by ISPs. Ketan Mayer-Patel (UNC) compared the two work presented by Michela and Nazanin -- one tries to bring structure into the overlay while the other tries to remove structure. Mohamed Hefeeda pointed out that increasing randomness in overlay topology seems to be counter intuitive. Nazanin explained that, without random links among peers, content might not flow to other ISPs.

Padmanabhan asked how else can P2P applications become location aware if ISP information is not available. The audience and presenters mentioned several alternatives, including Network Weather Service, CDN-based Relative Network Positioning, and the P4P project.

Padmanabhan also asked whether the overlays account for correlated churn, e.g, when a large number of peers join and leave simultaneously. Such churn, however, are not addressed by the papers.

Hassan Shojania presented their experience of implementing random network coding on the Apple iPhone and iPod Touch platform. The presentation overviewed network coding technologies and the challenges encountered in migrating the algorithm on a mobile device, with special attention to the lack of SIMD instruction set on the iPod. The presentation compared the coding efficiency over iPod Touch and iPhone, and presented the results of the feasibility study. The presentation also compared CPU-based and a GPU-based implementations.

Robin inquired about the effects of network coding on battery usage. Hassan replied that the interface to power management on the current iPhone is not exposed. He added that the CPU usage was, however, non linear. Hassan said that they were looking forward to the new processor in the forthcoming iPhone, which he believed had better SIMD instruction set support.

Padmanabhan started his presentation with the question ``What can one do in a room with 1000 cameras?'' The presentation then focused on the scale of data and the amount of processing required to extract some constructive information from data from these cameras. The core of the presentation was a runtime system called Sprout, used to identify and split parallel tasks across hundreds of machines with a final goal of achieving low latency for interactive applications with large data processing. The presentation focused on processing event/gesture recognition in videos in "interactive time" and identified three steps in the process: (i) identification of low-level features, (ii) matching low-level data against a training set to extract events of interest, (iii) aggregation of these results into appropriate events/gestures. Padmanabhan noted that the second step involves a sweep over the stream in both time and space, which can be parallelized (even within each frame), providing significant reduction in latency.

Klara Nahrstedt (UIUC) asked how close they were to achieving interactive time. Padmanabhan replied in short "very far" and said that the acquiring time from the cameras was huge since they used off-the-shelf systems. The presenter also added that the networks are not optimized as yet to handle such sudden bursts of packets from multiple synchronized cameras. Dongyan Xu then inquired if they had looked into any compiler techniques and generic code optimization techniques. Padmanabhan said that they had not done any low-level parallelization and had focused on parallelization at a higher level.

Philip W. Frey from IBM Research started with an assumption that "the network is no longer the bottleneck, but the memory to NIC bus is". The presentation focused on reducing the data fetch request processing time on server by removing the current 4-way copy (two DMA copies and two CPU copies) and two context switches required to send data to an incoming request. The authors proposed new protocol based on RDMA (Remote Direct Memory Access). The presentation showed the CPU usage in HTTP (with and without kernel sendfile() support) and RTP-based servers. They then presented their new client-driven protocol based on RDMA. Their impressive results showed a zero CPU load using the protocol. They also presented results showing lower context switches and lower interrupt rates. For live-streaming applications, direct memory transfer made their solution a zero-copy solution.

There are several clarifying questions from the audience. Philip explained that RDMA works over TCP and the address space used in the experiments is virtual. A follow-up question was on the DMA transfer that maybe required for virtual memory access. The speaker replied that such memory access is a one-time cost.

Damien Le Moal (Hitachi Development Laboratory, Japan) spoke about a new method of optimizing disk access in an effort to save power consumption. The presentation distinguished between normal data access and audio-visual (AV) data access targeted for set-top boxes, but general enough for use in normal computing systems. The idea is to have a separate file system for AV access, where each AV data request would be associated with a deadline it must meet. Thus, the system can hold all such requests in a queue and schedule them together in a more efficiently re-ordered fashion (according to seek distances). This increases the efficiency of the system by reducing the seek distance between two consecutive requests. The disk can also be spun down when the deadlines are far enough.

Mohamed Hefeeda inquired if scheduling these real time events in such a manner would adversely affect interactivity of multimedia application. Damien replied that as long as the deadlines are selected correctly, it will actually improve the interactivity since the event is now certain to be scheduled before its deadline expires.

Saurabh Ratti (University of Ottawa) presented, on behalf of the authors, a distributed algorithm for dynamic construction of an overlay topology that supports greedy routing in massively multi-user virtual environments. The idea is to partition the topology into two non-overlapping sets, where each set is updated in a separate phase using the other set as a reference. After explaining the theoretical background, Saurabh illustrated the algorithm steps and concluded with simulation results.

Dongyan Xu asked how to ensure the robustness of the algorithm against when nodes leave. Saurabh answered that the reliability would be guaranteed if TCP is used. If UDP was used, node leave would be detected in the next update circle. Robin suggested installing a scanner in each geometrical region for detecting the mobility of avatars. Saurabh argued that a decentralized solution is more preferable than centralized approach. Wei Tsang asked if there is any drawback of the algorithm. The response from Saurabh was that the resumption time presented in the simulation results is considerable. Although increasing the number of nodes has little effect on the resumption time, the baseline value of 1.1 seconds is not optimal, especially if the game has multiple updates within a second.

John Miller (MSR and University of Cambridge) proposed the use of Pairwise Random protocol (PRP), which uses secure coin flipping to fairly determine the resolution of competition between two avatars in distributed virtual environments. He started his talk with an introduction to some related DVE security researches and explained subsequently the three-way message exchange included in PRP protocol. Performance analysis of three different variations of PRP was presented. John concluded that the choice of the PRP variations was a trade-off between overhead and security compromise.

In response to Dongyan Xu's question about the reliability of the proposed approach, John said that the PRP protocol assumed reliable transmission between the avatars. Wei Tsang asked about the possibility of extending the PRP protocol to more than two parties. The response was that it was possible to decompose the message exchanges between multiple parties, but the extension would be challenging due to the increased complexity and performance overhead. Ardalan Kangarlou (Purdue University) asked whether there was any mechanism that could validate the hash value used in PRP. John answered the exchanged message itself can be used to prove the correctness of the delivered information.

Ke Liang presented an algorithm that constructs and adjusts multicast trees for spatialized audio streaming in a peer-to-peer overlay network. The algorithm has two objectives: Maximize total number of audio receivers, subject to nodes' bandwidth limits; while minimizing the average latency of those receivers. Ke interpreted their solution to the problem as achieving a compromise between the two objectives via maximizing the number of receiver that have the minimum latency. Extending from their previous work, Ke proposed the cross-tree adjusting (CTA) algorithm that can re-allocate the upload bandwidth incrementally for nodes with bandwidth conflicts in all existing multicast trees. Simulation results showed that CTA can achieve good performance in terms of fairness and the total number of successful receivers that have low latency.

Klara Nahrstedt asked if the cocktail effect had been considered in the proposed algorithm. The cocktail party effect describes the human's ability to focus one's listening attention on a single talker among a mixture of conversations and background noises. Ke agreed on that the cocktail party effect could be something for further investigation, although it was not covered by their current work. A suggestion to their future work from Klara was that the distance between avatars in the virtual world could be taken into account when constructing the multicast trees.

Mohamed Hefeeda presented their work on the problem of securing the delivery of scalable video streams so that receivers can ensure the authenticity (originality and integrity) of the video. Their focus in the paper is on recent scalable video coding techniques, e.g., H.264/SVC, which can provide three scalability types at the same time: temporal, spatial, and quality (or PSNR). Mohamed presented an efficient authentication scheme that accounts for the full scalability of video streams, enabling verification of all possible substreams that can be extracted and decoded from the original stream.

Dick Bulterman (CWI) wondered if there is any standard fee needed to be charged to scalable video coding techniques. The presenter responded that they are using the open source standard and there are a lot of companies (e.g. Google) that have already use the standard and the reference source code.

Tony Thomas (Nanyang Technological University) presented a joint watermarking and cryptographic mechanisms for securely delivering multimedia content on multiparty multilevel digital rights management (DRM) architecture, where content are delivered by an owner to a consumer through several levels of distributors. The presenter emphasised that license is more important than content since the content will be encrypted before delivery from the owner or distributors. The authors proposed a mechanism that minimizes the possible degradation of the quality of a content due to embedding of watermark signals. Furthermore, in case the owner or a distributor finds an unauthorized copy, they can identify the traitor with the help of a judge.

Philip Frey (IBM Research GmbH) asked what is the advantage of multi-layer design. Tony answered that the multimedia content is distributed by multiple distributors which will encrypt the contents before distribution. From the client's point of view, it could track the distributor according to the content it received. Alexander Eichhorn (Simula Research) asked why multiple layers of license servers are used. The presenter responded that it is due to business issues. Since clients may be geographically distributed, multiple distributors that may reencrypt the content are used. Thus, client need to request different licenses from different license servers.

In this talk, the presenter presented results of experimental work using machine learning techniques (decision tree) to rapidly identify Skype traffic. They use a number of feature classes to classify IP flows as being Skype or non-Skype flows. The feature classes they found most effective were the interarrival times between IP packets frequently occurring IP packet lengths less than 80 bytes. Their results show that using three feature classes in a single classifier provided 98 percent precision and 99 percent recall when using a window duration of 5 seconds or more. The presenter emphasizes that their classifiers do not rely on observing any particular part of a flow. They also report on the performance of classifiers built using combinations of two of these features and of each feature in isolation.

Amir Hassan Rasti Ekbatani (University of Oregon) asked the presenter how to detect audio flows. The presenter answered that their approach works on small sequence of any part of flow. Since they make use of techniques of machine learning, they can identify Skype traffic reliably with only a few seconds of traffic, whose characteristics are extracted from short sliding windows and used to train a classifier. Amir also asked if the authors tried using larger packet size, to which Philip answered that they only used SVOPC codec so far with packet size less than 80 bytes. Michela asked what happened when Skype traffic has both audio and video data (i.e. packet size is increased). The presenter answered they still do not have a good solution since their classifier is trained from pure audio flows.

In VoIP applications, buffer size needs to be carefully chosen to trade-off between audio quality and delay. Finding optimal buffer size is challenging because many factors are involved. According to their experimental results, real-life VoIP applications, such as Skype, MSN, and gTalk, do not adjust the buffer size well. They proposed a regression-based lightweight algorithm to compute the optimal buffer size in real time.

Wei Tsang pointed out that some VoIP applications may apply time compression to drop the silent period. He asked whether time compression might affect buffer size estimation. Kuan-Ta Chen said that estimating the buffer size according to time shifting is the only thing that can be done currently. Finding out whether time compression is used by reverse engineering may work better but needs large number of experiments. Mohamed asked which factor affects the user experience most as many factors exist. Kuan-Ta answered that it is still an open question and his personal opinion is that the codec may be the most important factor.

In scalable video streaming, the granularity of bit rate adaptation with CGS is limited, while using MGS brings high signal overhead. Pengpeng Ni proposed that by switching between layers quickly, arbitrary bandwidth could be achieved without high overhead. To understand and compare the user experience when different switching patterns are used, the authors conducted a user study. The result shows that frequent layer switching is effective and could be better than quality downscaling, with the switching period as a crucial design parameter.

Damien Le Moal asked whether the nominal frame rate is considered. Pengpeng answered that only 25 and 12 fps is used in user study as covering different frame rate variations in the user study is time-consuming. Saurabh asked whether the consistency of the user response is considered. Pengpeng answered that during binary measurement, two sequence may appear twice with different order (A, B and B, A), so it could be used to test the consistency of user response. Padmanabhan asked whether experiments are done on CRT since LCD, which has lower refresh rate, may reduce the negative effect of low frame rate. Pengpeng pointed out since the field study is done on iPhone, only LCD is used. Mohamed asked whether the lighting condition is considered and whether the experiments are done following the ITU regulation. According to Pengpeng, they find field study more interesting than experiments in a controlled lab since the former is closer to the real world usage.

Ishan highlighted that scheduling at intermediate nodes is essential in soft real-time applications. After he briefly introduced the traditional IntServ and DiffServ and their disvantages, Ishan proposed a new method that compares the estimated transmission time and deadline in each node and schedules the packet with shortest per-hop time first. The main advantage of this method is (i) bursts are better handled, (ii) packets cannot meet the deadline are discarded to avoid wasting bandwidth, and (iii) the server is stateless.

Alexander asked which router is used in the implementation since this method requires clock synchronization and other computation. Ishan answered that currently only Linux machines with IP forwarding enabled are used, but they plan to use Cisco routers in the future. Padmanabhan asked about misbehaving node that reports false deadline requirement to gain advantage. Ishan argued that packets that cannot meet deadline will be dropped, so a misbehavior flow may be dropped to benefit other flows. John Miller asked how to obtain the value of TTL and deadline. Ishan answered that TTL value can be measured by nodes in the first round trip transmission and the deadline is decided according to the requirement of the application.

Robin opened his talk by emphasizing on the difference between Internet TV and IPTV. The former is applied to the best effort delivery of video content over a broadband connection and the latter is defined by ITU-T to provide security, reliability, and interactivity. He showed AT&T's architecture for IPTV, where optical media connects the video source to the DSLAM switch, which are connected to models using high bandwidth DSL links. The uplink bandwidth of a DSLAM switch is 1 Gbps and thus only supports a small number of IPTV users concurrently. The speaker then explained their proposed peer-assisted system for IPTV, Zebroid, where popular video-on-demand content is pre-striped on user set-top boxes (STB) during idle hours. The STBs will then form a peer-assisted streaming overlay to deliver the VoD content to each interested user during the peak hours. Zebroid also tries to predict the probability of a node failure which happens when an STB is turned off and provision a level of redundancy accordingly. The simulation-based evaluation shows that their proposed peer-assisted delivery is only possible for the 8Mbps and 12 Mbps neighborhoods and only when the number of requesting peers is at most 8.

Remo Meier started by talking about the challenges introduced by the P2P application model, including peer heterogeneity, robustness against randomness and selfishness, as well as system's fairness among peers. The presenter mentioned the features of a hypercubic overlay including simplicity of construction and maintenance as well as neighbor selection flexibility. Remo then presented the proposed P2P streaming mechanism in which the peers form multiple hypercubic trees and push the content blocks down the tree to reach a quarter of the peers. In the next phase, with a BitTorrent-like pull-based mechanism, the content is distributed among all peers. Using simulation-based evaluation, Meier showed that their proposed system successfully limits free riders. Peers who share too few packets often suffer from underflows. He suggested that by using layered audio/video codecs, it is possible to accommodate heterogenenous peer groups.

Lisong Xu opened his talk by positing that a common problem for P2P streaming was the ``best-effort'' quality. Due to user dynamics and system complexity, achieving a guaranteed quality is hard. The speaker defined the paper's goal as providing statistical guarantee on the overall upload bandwidth of a P2P streaming system using admission control. The paper uses a queueing model used based on heterogeneous upload bandwidth and includes user dynamics as well as admission control, and defines the statistical guarantee problem as deciding whether to admit a new peer to maintain the probability of having enough upload bandwidth. Toward this end, they show that a dynamic admission control algorithm that exploit state information leads to lower rejection rate with the same level of guarantee. The speaker also explore admission control's sensitivity to user lifetime distribution and user arrival process. and concluded that there is a fundamental tradeoff between user rejection rate and user behavior insensitivity.

The session continued with a question/answer panel including all three speakers.

Mohamed asked the last speaker where the admission control algorithms were to be executed. Lisong pointed out that their paper had assumed a central entity responsible for admission control. Mohamed then asked the first speaker whether in Zebroid provides multiple encoding rates or scalable oding for the same TV channel to accommodate users who want to receive HD or SD versions of each channel. Robin Chen was not sure whether multiple encoding rates were provided, yet he pointed out that in future such schemes would be necessary to provide IPTV service to mobile devices.

Ardalan asked Remo about the possibility of unbalanced delivery tree. Remo responded that due to randomness of IDs, unbalanced tree is rare, however, their algorithm periodically ensure that the tree is balance.

Michela raised the concern of energy consumption at the set-top boxes in Zebroid during the idle hours when the boxes would be otherwise off or stand-by. Robin Chen explained that the set-top boxes have the potential to be the next battlefield for a variety of applications similar to cell phones and could stay on most of the time.

Hassan asked Robin about the mention of erasure code in his presentation. Robin clarified that erasure code is necessary to compensate for the peers that are not up during the content push time or at the peak hour. However, such mechanism was not implemented into Zebroid yet.

Saurabh asked Robin about the level of peer outage considered. Robin explained that they assume 80% of the peers are up at all times, and this rate is based on measurement data.

Mohamed asked Robin if data pushed to each client has any correlation with the usage history of that user. Chen responded that they plan to include a mechanism to predict the usage of each user and consider that prediction in pushing the data and then refining the predictions gradually with the user's choices. Discussion ensued about the privacy issues on this approach.

Philip asked Remo about ID conflict between clients since IDs are assigned randomly. Remo Meier explains that although the probability is low, they had proposed a mechanism to ensure that does not happen in another paper.

Dongyan then asked the panelists how they assessed their proposed systems' robustness. Robin mentioned that they could configure the system to ensure that the user had no access to the set-top box, while Remo pointed out that their approach designed each peer's behavior to be as greedy as possible, therefore they would not expect any manipulation to change the code to make it more selfish.