PeerGraph: A Distributed Data Structure for Peer-to-peer Streaming

Transcription

1 PeerGraph: A Distributed Data Structure for Peer-to-peer Streaming Ali Şaman Tosun and Turgay Korkmaz Department of Computer Science Uniersity of Texas at San Antonio San Antonio, TX 79 Abstract Video streaming oer peer-to-peer networks has receied significant attention recently. Accordingly, much work is done on ideo streaming under the assumption that all the peers hae the whole moie. Howeer, it is likely that many peers may only store a fraction of the moie due to their resource constraints. Therefore, there is a need to deelop appropriate mechanisms to facilitate efficient streaming when each user stores a fraction of the moie. With this in mind, we propose a distributed data structure called PeerGraph in which each peer determines other peers haing the preious or subsequent fractions of the moie and maintains pointers to these peers. In our model, a peer has full control on how many segments it wants to store. Howeer, we require it to store consecutie segments. In order to stream, a client first finds a peer that contains the first segment and then uses the proposed PeerGraph to locate subsequent segments. We use extensie simulations to inestigate the effect of arious parameters. Simulation results support feasibility of the proposed scheme. I. Introduction For many decades, ideo has been an important media for communications and entertainment. Attempts to display media on computers date back to the early days of computing. Howeer, little progress was made for decades, primarily due to the high cost and limited capabilities of computer hardware. Neertheless, recent adances in computer networking combined with powerful home computers and modern operating systems made multimedia systems practical and affordable for ordinary consumers. In general, multimedia systems can be found three different forms []: stand alone systems, systems connected ia dedicated links, and systems connected ia the public Internet. While the popularity of the first two forms has increased at a rapid rate, the last form could not be widely used due to the shortcomings of the current best-effort Internet such as unknown and time-arying bandwidth, delay, losses and difficulty in fair sharing of the network resources. To oercome some of these shortcomings without changing the Internet, researchers hae been extensiely inestigating arious multimedia streaming schemes. Seeral studies (e.g., [], [], []) focus on streaming of pre-encoded ideo which is the most common method today. Some focus on encoding and streaming ideo for applications such as ideoconferencing [], [], [], [], []. In general, streaming schemes try to address seeral key issues by using different mechanisms such as controlling error including retransmissions, added redundancy using forward error correction (FEC) [7], [], [], error concealment [], [] and error resilient ideo coding [7]. To improe streaming oer best-effort networks, researchers also proposed seeral mechanisms including the use of content deliery networks [], [], [], caching [7], [], [], patching [7], [] and transcoding []. Next frontier in streaming is peer-to-peer networks which introduces additional challenges such as streaming from multiple nodes and selecting a set of peers. In [], an optimization based framework is proposed to retriee packets from the node which minimizes loss and delay. Using FEC together with rate allocation [] in distributed streaming reduces probability of packet loss in bursty loss enironments caused by network congestion resulting in higher isual quality for the streamed ideo. PeerCast [] streams lie media using an oerlay tree formed by clients and CoopNet [] proposes a mechanisms for cooperation of clients to distribute streaming ideo when serer is oerloaded. A peer-to-peer media streaming model with an optimal media data assignment algorithm and a differentiated admission control protocol is proposed []. Hybrid CDN, peer-to-peer media distribution results in efficient

2 use of CDN resources []. Layered peer-to-peer streaming handles asynchrony of user requests and heterogeneity of peer network bandwidth []. Administratie organization of peers results in reduced control oerhead in media streaming []. Promise peer-to-peer system [9] supports peer lookup, peer-based aggregated streaming and dynamic adaptations when network and peer conditions change. Peer-to-peer streaming in which each peer stores partial ideo is proposed in [9]. This approach uses on demand approach of segment selection. In this paper, we perform segment selection using a distributed data structure. All of the aboe techniques assumes that the whole moie is stored at all the peers. Accordingly, they focused on how to choose the peers based on their delay, loss and outgoing bandwidth. Howeer, if peers store partial ideo, a whole new set of challenges are introduced including the following: How can a client determine whether a gien set of peers are enough to stream the ideo? Gien space for segments, how can a peer determine the segments that it stores? Should a peer cooperate with other peers to determine which fragments it stores? PSfrag replacements How much control information needs to be exchanged to determine the supplying peers when a client requests a moie? In this paper, we propose a distributed data structure that can be used for peer-to-peer streaming when each peer stores a fraction of the ideo. Proposed data structure organizes peers thereby reducing message oerhead to find peers for streaming. Proposed scheme allows indiidual nodes to store as many segments as they want and store the segments they choose (with some restrictions). Simulation results support feasibility of proposed data structure. Rest of the paper is organized as follows: we describe the system model in Section II, join and leae algorithms in Section III. Section IV presents simulation results. Finally, I we conclude this paper in Section V. II. System Model We describe the proposed data structure for a single moie, a separate data structure needs to be constructed for each moie. We assume that the moie is diided into equal size segments and segment is denoted by. A client stores consecutie segments of the moie after he/she iews the moie. The number of segments stored by the client depends on storage space and bandwidth aailable at the client and is determined by the client. If a client is able to store segments, the moie is diided into blocks each haing segments and the client picks one of the blocks randomly. Note that the last block may hae less than segments in it if is & nonzero. The segments in "! #$% by the following formula. ( "! # %*),+.- / $.- / $ ' are gien 7 ' 7*)9 For example, possible blocks for );: and <)>= are :A?.B C, A? and =A. In this case last block has a single segment. The client stores consecutie segments of the selected block. Storing consecutie segments has two main adantages: it reduces oerhead of the data structure and uses a good peer as long as possible. Selection of segments based on the blocks guarantees that each segment appears in the system approximately equal number of times. To enable faster access to the prefix of the moie, probability of selecting the first block can be increased resulting in larger number of nodes which has the prefix. 7 Fig.. Example Data Structure. Forward Ptr Backward Ptr Reerse Ptr Suppose each peer independently selected a range of segments and stored them. Now, the key question is how to stream the whole moie. For this, we propose a distributed data structure called PeerGraph in which each peer determines other peers haing the preious or subsequent segments of the moie and maintains pointers to these peers. Figure illustrates a PeerGraph for D)>= segments and EF)HG peers. A streaming session starts from one of the nodes storing (e.g., nodes,, in Figure ) and follows one of many forward links at each step to reach a node storing (e.g., one of nodes,, in Figure ). The goal of haing multiple links is to let the streaming node choose one good peer based on loss and bandwidth. The streaming node can probe all the potential links by computing loss rate and aailable bandwidth and pick one based on the results. The tests can be interleaed with streaming to preent gaps in streaming experience. Nodes with content can join and leae the system. Therefore, join and leae operations should be efficient. We now formally define the proposed (PeerGraph) data structure. The notation used throughout the paper is gien in Table I. In addition, we use the following terms: forward pointer: A forward pointer from node J to node K means that next segment is stored at node K.

3 L M w ~ ~ \ \ C % ¾ ¾ w C PSfrag replacements Notation Meaning Number of Segments Number of Nodes in PeerGraph NPO QSRST segment U'V Q V L NXWZY[ \ Maximum segment id stored at node N W O_^ ] \ Minimum segment id stored at node ` ] \ ` \ba NcWZY[ \ d NcW O_^ f*gih \ ^ R \ e U \ Number of forward pointers at node j*gih \ ^ R ] \ Number of backward pointers at node f'k ] \l Q$m Q RST forward pointer at node ]nu'v Q V f gih \ ^ R j'k \ l QSm Q RST backward pointer at node ]nu'v Q V j'goh \ ^ R pzq \ reerse pointer at node ] TABLE I. Notation Used. rtsu yx{z 'r } 'rtsu More formally ~. In this case K is called a forward neighbor of node J. backward pointer: A backward pointer from node J to node K means that preious segment is stored at node K r }. More formally yx{z 'r _ 'rtsu ~. In this case K is called a backward neighbor of node J. reerse pointer: A reerse pointer from node J to node K means that node J is a forward neighbor of node K. Reerse pointer is used to guarantee that each node I has a path to itself from a node with segment. Using aboe notation, we can define PeerGraph data structure as follows. Definition: : A PeerGraph with parameters Zƒ and ƒ is a distributed data structure for a moie with segments with the following requirements Each node (except nodes with segment ) has z forward pointers each pointing to a different node. z Each node (except nodes with segment ) has backward pointers each pointing to a different node. Each node (except nodes with segment ) has a single reerse pointer. In essence, PeerGraph consists of seeral nodes. Each node J z r } rtsu maintains the segments in the range and tries to determine at least and at most forward pointers and backward pointers. For each forward pointer #$% the number of segments and index of maximum segment at node # ˆ% is stored at node J to efficiently perform the join operation. Similarly, for each backward pointer #$% the number of segments and index of minimum segment at node #$% is stored at each node J. In addition, a single reerse pointer Š is stored at each node. Reerse pointer Š is used to guarantee that there is at least one path from a node with segment to the current node J. For example, Figure shows the node structure of node 7 in Figure. Node 7 has forward pointers pointing to nodes and Œ. Node stores segments maximum segment being Œ and node stores segment maximum segment being. #$% - } 'rtsu - } Z # ˆ% Ž - } r } - } Fig.. Node Structure of node 7. Node 7 has backward pointers pointing to nodes and *. Node stores segments minimum segment being and node stores segments minimum segment being. Reerse pointer Šy points to node. Based on the definiton of PeerGraph data structure space requirement. per node is : # w :S }P "% w III. PeerGraph Algorithms A. Join Operation When a node J with segments of r } Š 'rtsu joins the system, forward and backward pointers should be correctly set up according to the system parameter. To set up forward pointers node J needs to find between and rtsu 'rtsu nodes with segment (unless )H ). To set up backward pointers node J needs to find between and r } nodes with segment r _ (unless ) ). Join algorithm is gien in Algorithm. First, new node J * searches for a node that contains the first segment using the function FindAStartNode. Finding a start node is the only additional support needed from the network, eerything else is done using the PeerGraph structure. Algorithm š œ#$j ž : Ÿ 9 $c «ª (â n "±$² ³tµ : while W O ^ \ ¹»º ¼ ½ W O_^ ¾ WZY[ À do : ŸÁ c " Ã Ä ² : end while : {Ædz â ² Ȫ (.É ±ÅŸ W O_^ \ µ : ª Êb ËÌPÄ ² $cª É ±SÍ µ 7: ª Â Î É (² $Èâ P±SÍ µ : while \ WZY[ ÏÐ º ¼ ½ W O ^ ¾ WZY[ À do 9: ŸÁ c " Ã Ä ² : end while WY[ \ ÏÐ µ : {Ædz â ² $cª É ±ÅŸ : â Ä ² $cª É ±SÍ µ Once a start node is selected, the node J searches for its position in PeerGraph by following a randomly selected forward pointer using the function RandomForwardPointer. When a node (say Ñ r } ) that contains is located, the backward pointers need to be setup. For this the algorithm finds all the backward neighbors of forward neighbors of Ñ and all the forward neighbors of

4 % w Ñ Ñ èæ ê ) ê ) backward neighbors of Ñ. As a result, a candidate set is constructed. Next, the pointers in the candidate set are I checked to see if they hae the desired segment and if they do they are placed in the set Š. Then nodes are picked from the result set. Next step is to set the reerse pointer. In essence, this algorithm uses the same candidate set found aboe for setting backward pointers, and ask if any of these nodes can add a forward pointer to node J. Finally, node J searches for a node that contains rtsu. When such a node Ñ is located, the forward pointers need to be setup. First a candidate set of pointers is constructed and nodes are picked from the result set. Complexity of node join operation is Ò# w o% since in the worst case each node can store a single segment causing traersal of nodes and setting up forward and backward pointers requires Ò#oo% cost. B. Leae Operation A node J with segments r _ r s(u leaes the system after it notifies its forward neighbors that it is leaing the system. Leae algorithm is gien in Algorithm. Forward pointers are notified because they Algorithm Ó Ô.ÕÖKÔ#$J ž : for all ª gih \ ^ R : n (ª o ÖØP n "±$ t²y±s iµˆµ : end for \ do might hae a reerse pointer pointing to node J. Consider node 7 in Figure as an example. Nodes and hae reerse pointers pointing to node 7. When node 7 leaes the system all the forward neighbors (node and 7 in this case) are notified. Nodes and can run Ô.ÙŠyÔ.KÖÔ.Ú Û.Ô ž œ ÙÔ.Ú routine to establish new reerse pointers. In Figure nodes and hae forward pointers pointing to node 7 and when node 7 leaes the system the pointers will not be alid. In PeerGraph each node periodically erifies that its neighbors are alie using short Ü Ú"Ô Ýn J«Ü&$ KÖÔ messages. Using this mechanism nodes and can detect that node 7 has left the system and establish new neighbors is necessary. IV. Simulation PeerGraph data structure is simulated using discrete eent based simulation library CSIM []. Each node is implemented as a separate process in CSIM. We first create desired number of nodes and randomly select the range of segments stored at each node. For these, we hae a parameter yþ ÕÖß which bounds the fraction of moie stored at a node. In simulations yþ ÕÖß aries from. yþ and ÕÖß, each node J determines? P ( by selecting a random number from the set Dà? P ( yþ ÕÖß A and then selects a random block from the set PA. We conducted seeral simulations to understand the impact of arious parameters on the performance. Due to space limitation, we focus on a small subset of interesting simulation results. yþ In Figure (a), we demonstrate the impact of ÕÖß to.9 with increments of.. Gien and E on the number of nodes used to successfully complete streaming requests. When the number of nodes that store a moie is small, in most cases, it is not possible to complete a streaming session using a PeerGraph data structure. With the increase in number of nodes, we may complete a streaming session by combining the segments stored at many nodes, increasing the number of hops inoled in successful streaming. Simulation results show that this is the case until a threshold. After the threshold, the number of hops does not increase with the yþ addition of more nodes. Instead, we see that increasing ÕÖß rather than adding new nodes is a better strategy to reduce the number of hops in streaming. yþ In Figure (a), we look at the impact of Õ@ß and node dynamics on the number of hops inoled in successfully complete streaming. As expected, the number of hops decreases as the number of segments stored at each node increases. Howeer, the fraction of nodes that leae the PeerGraph does not affect the number of hops much since the nodes dynamically call join procedure when the conditions in Definition are not satisfied. Next, we consider the factors impacting the number of forward links. When forward neighbor of a node leaes, the number of forward pointers decreases. When a new node J adds a reerse pointer to node K, it forces node K to add a new forward neighbor to node J. To accommodate such scenarios, if possible, we add between -}á Sâ and forward pointers for a new node using a probabilistic scheme. When the set of candidates Ü for forward neighbors is computed, the element in position is added as forward neighbor with probability Ñ, which is defined as follows. y ç - $/«â 7 )åäæ -}á / Sâ -}á Sâ 7é -}á Sâ ƒ is linear between -ëá ZSâ and Clearly, Ñ. The alue of decreases from to between-}áthese Sâ alues and Ñ when *); and Ñ when '). In Figure (b) we demonstrate the impact of number of nodes and yþ ÕÖß on the number of forward links. As the number of nodes increase, each nodes has a chance of increasing the the number of forward links. In Figure (b) we consider the impact of node dynamics on the number forward links. As nodes leae, the number

5 Number of Hops % % % N=, L=, K= Number of FWD 7 N=, L=, K= % % % Number of BCK N=, L=, K= % % % (a) Number of Hops (b) Number of Forward Pointers (c) Number of Backward Pointers Fig.. Impact of SPMax on System Parameters Number of Hops N=, M=, L=, K= no nodes leae % leaes % leaes % leaes Number of FWD 7.. N=, M=, L=, K= no nodes leae. % leaes % leaes % leaes Number of BCK.. N=, M=, L=, K= no nodes leae % leaes % leaes % leaes (a) Number of Hops (b) Number of Forward Pointers (c) Number of Backward Pointers Fig.. Impact of Node Dynamics on System Parameters of forward pointers decreases since decrease in number of forward pointers does not automatically trigger reconfiguration because we probabilistically maintain more than links. When adding backward pointers, we use the same probabilistic approach used aboe for forward pointers. Figure (c) and Figure (c) show the number of backward pointers under the same scenario considered for forward pointers. In general, we see the same trend as in forward pointers. That is () number of backward links increase as the number of nodes increase and stays constant after a threshold, () number of backward pointers decreases as nodes leae the system but not much affected. As nodes join and leae the system, some nodes may end up haing less than backward or forward pointers and a path from the node to a node with segment may not exist. This causes an incomplete streaming session. We measure this effect using success rate metric. Success rate for different scenarios is gien in Figure (a) and Figure. As nodes leae the system, success rate decreases and storing more segments at a node results in better success rate. One of the design goals of PeerGraph data structure is to determine replication decisions independently, yet guarantee some balance of segments among the actie Sucess rate of Complete Streaming.9.9. N=, M=, L=, K=. no nodes leae.7 % leaes % leaes % leaes Fig.. Success rate of complete streaming nodes. We proide the segment distribution here to show this fact. Figure (b) shows for each segment the fraction of actie nodes containing the segment. As nodes leae the data structure the fraction decreases. Howeer, it is still balanced. PeerGraph data structure does not guarantee that the number of incoming forward pointers to nodes will be balanced. We next inestigate this using simulation results. Assume that node J stores / only the segment. Only nodes that store segment can hae a forward pointer to node J by the definition of forward pointer. Assume another node K that stores the / íì segments. ì All / the nodes that store segments through segment can hae a forward pointer to node K. So, the more the

6 Sucess rate of Complete Streaming N=, L=, K=.. % %. %. Prob of being in an actie node N=, L=, K=, M=. % %. % Segment id Ag number of FWD links N=, L=, K=, M= % % % Size of the stored segment range (a) Success Rate of Complete Streaming (b) N@O Fraction of Actie Nodes Containing Fig.. Impact of on System Parameters (c) Ag. Number of Incoming Forward Pointers number of segments stored is, the more potential there is for incoming forward pointers. The data from simulation results is gien in Figure (c). x axis denotes the number of segments stored at the node and y axis shows the aerage number of incoming forward pointers. The increasing trend can be obsered in the figure as the number of segments stored is increased. This trend is based on the number of segments stored at a node and independent of the number of actie nodes. V. Conclusion and Future Work We propose a distributed data structure called Peer- Graph for streaming in peer-to-peer networks. Proposed data structure allows nodes to store the ideo partially while enabling efficient streaming. Replication decision is made independently by each node in a way that guarantees balanced segment distribution among actie nodes. We inestigate the proposed scheme using extensie simulations. Simulation results show the feasibility of the approach. References [] J.G. Apostolopoulos, T. Wong, W. Tan, and S.J. Wee. On multiple description streaming with content deliery networks. In IEEE INFOCOM, June. [] Songqing Chen, Bo Shen, Susie Wee, and Xiaodong Zhang. Designs of high quality streaming proxy systems. In IEEE INFOCOM, July. [] Philippe Cuetos and Keith Ross. Optimal streaming of layered ideo: joint scheduling and error concealment. In UU RST ACM Multimedia, Noember. [] Yi Cui and Klara Nahrstedt. Layered peer-to-peer streaming. In NOSSDAV. [] Philippe de Cuetos and Keith Ross W. Unified framework for optimal ideo streaming. In IEEE INFOCOM, July. [] H. Deshpande, M Baea, and H. Garcia-Mollina. Streaming lie media oer peers. Technical report, Stanford Database Group Technical Report (-). [7] Lixin Gao, Zhi-Li Zhang, and Don Towsley. Proxy-assisted techniques for deliering continuous multimedia streams. In IEEE/ACM Transactions on Networking, December. [] Meng Guo and Mostafa Ammar. Scalable lie ideo streaming to cooperatie clients using time shifting and ideo patching. In IEEE INFOCOM, July. [9] Mohamed Hefeeda, Ahsan Habib, Boyan Bote, Dongyan Xu, and Bharat Bhargaa. Promise: Peer-to-peer media streaming using collectcast. In ACM Multimedia. [] Zhijun Lei and Nicolas Georganas. Video processing and transformation: Rate adaptation transcoding for precoded ideo streams. In Uî RST ACM Multimedia, December. [] Jiangchuan Liu, Xiaowen Chu, and Jianliang Xu. Proxy cache management for fine-grained scalable ideo streaming. In IEEE INFOCOM, July. [] Mesquite Software, Inc. CSIM 9 Simulation Engine (c ersion). [] T. P. Nguyen and A Zakhor. Distributed ideo stresming oer internet. In SPIE/ACM MMCN. [] Thinh Nguyen and Aideh Zakhor. Distributed ideo streaming with forward error correction. In Packetideo Workshop. [] V.N. Padmanabhan, H. Wang, P. Chou, and K. Sripanidkulchai. Distributed streaming media content using cooperatie networking. In NOSSDAV. [] S. V. Raghaan and S. K. Tripathi. Networked Multimedia Systems: Concepts, Architecture, and Design. Prentice Hall, Inc., 99. [7] K. Stuhlmuller, N. Farber, M. Link, and B. Girod. Analysis of ideo transmission oer lossy channels. In IEEE J.Select Areas Commun., June. [] Duc Tran, Kien Hua, and Tai Do. A peer-to-peer architecture for media streaming. Journal in Selected Areas in Communications, Special Issue on Adances in Serice Oerlay Networks, ():, January. [9] Sailaja Uppalapati and Ali Şaman Tosun. Partial ideo replication for peer-to-peer streaming. In ï RëT International Conference on Management of Multimedia Networks and Serices, pages,. [] Chitra Venkatramani, Oliier Verscheure, Pascal Frossard, and Kang-Won Lee. Optimal proxy management for multimedia streaming in content distribution networks. In Uð RëT NOSSDAV, May. [] Zhiheng Wang, Sujata Banerjee, and Sugih Jamin. Studying streaming ideo quality: from an application point of iew. In UU RëT ACM Multimedia, Noember. [] Min Wu, Su-Jun Ma, and Wei Shu. Scheduled ideo deliery for scalable on-demand serice. In Uð RST NOSSDAV, May. [] Dongya Xu, Heung-Keung Chai, Rosenberg Catherine, and Sunil Kulkarni. Analysis of a hybrid architecture for cost-effectie streaming media. In SPIE/ACM Conference on Multimedia Computing and Networking (MMCN ). [] Dongyan Xu, Mohamed Hefeeda, Susanne Hambrusch, and Bharat Bhargaa. On peer-to-peer media streaming. In IEEE International Conference on Distributed Computing Systems (ICDCS ), July. [] Haim Zlatokrilo and Hanoch Ley. Packet dispersion and quality of oice oer ip applications in ip networks. In IEEE INFOCOM, July.