| 摘要: | 有一個在網際網路上廣播熱門影片的方法是先將影片切割成小的區塊, 然後重覆在多個頻道播放. 由於這種方法可以讓使用者分享頻寬, 所以可以在不犠牲等待時間的情況下, 節省頻寬. 按區塊的播放方式, 這種廣播可分為兩種: 每個頻道傳送一個區塊及每個頻道傳送多個區塊. 本論文主要討論基於後者的廣播法, 其中兩個代表性的方法為 Recursive Frequency-Splitting broadcasting (RFS) 及 Fixed-Delay Pagoda Broadcasting (FDPB), 其均可提供非常小的等待時間. 我們提出區塊補償與非同步下載及播放策略讓 RFS 及 FDPB 可以用較平滑的頻寬傳送 VBR 影片. 區塊補償是動態播放影片後面的區塊使得所需頻寬較平穏. 而非同步下載及播放是要求使用者必須下載完區塊後才能播放, 如此讓影片播放才能連續. 利用這兩個策略, 我們發展 Smooth RFS (SRFS) 及 Smooth FDPB (SFDPB), 可以傳送 VBR 影片並降低所需頻寬峰值及變異量, 我們也以數學分析其邊界值. 此外本研究以一部影片測試其效能, 實驗結果顯示 SRFS 比 RFS 有較低的頻寬峰值及變異量, 唯一的代價是稍高的平均使用頻寬. 與 Smooth Fast Broadcasting (SFB) 相比, SRFS 也有較低的頻寬峰值及平均使用頻寬. 當給定頻寬時, SRFS 的阻塞率遠比 RFS 及 SFB 小. 對 SFDB, 我們也得到同樣的實驗結果. 除此之外, 本論文也探討另一個重要的問題-降低使用者暫存器需求. 我們提出一個策略要求使用者儘量延遲區塊的下載, 並證明此策略可讓使用者儲存最少的區塊. 本文進一步提出一個反向排列區塊的策略, 其以由大到小的方式排列頻道中的每個區塊. 利用這兩個策略, 我們設計 Reverse Fast Broadcasting (RFB) 廣播法, 最大只需暫存25%的影片. 與 Fast Broadcasting, New Pagoda Broadcasting 及 RFS 相比, RFB 節省客戶端暫存需求50%, 33% 及 33%. 我們擴充此法提出 RFB-n, 可提供伺服器端頻寬及客戶端暫存需求的調節. 在額外給三個頻道的情況下, RFB-3 只需暫存10%的影片. 我們進一步將這兩個策略用於其他廣播法, 如 Greedy Disk-conserving Broadcasting 及 Pyramid Broadcasting, 發現可以節省客戶端暫存需求30%至75%. 最後, 我們也討論所提出策略在節省客戶端暫存的限制. One way to broadcast a popular video is to partition the video into segments, which are broadcasted on several streams periodically. The approach lets multiple users share streams; thus, the stress on the scarce bandwidth can be alleviated without sacrificing viewers’ waiting time. The segment broadcasting can be categorized into two types: one segment per stream and multiple segments per stream. Two representative schemes based on multiple segments per stream are the recursive frequency-splitting (RFS) broadcasting and the fixed-delay pagoda broadcasting (FDPB), which both obtain very small waiting time in the literature. We propose the approaches of segment patching and asynchronous downloading-playing to enable the RFS and the FDPB schemes to broadcast VBR-encoded videos smoothly. Segment patching dynamically arranges the video segments numbered latter to smooth bandwidth requirements. Asynchronous downloading-playing ensures that a client can play a VBR video continuously. With the approaches, the enhanced schemes, named the smooth RFS (SRFS) and the smooth FDPB (SFDPB), reduce peak bandwidth consumption and variance during distributing VBR videos. In addition, we analyze the bounds on the peak and the variance. A simulation was conducted to evaluate both the schemes. The results indicate that the SRFS outperforms the RFS scheme remarkably on reducing the bandwidth peak and the variance. The overhead is slightly higher average bandwidth. In comparison with the smooth fast broadcasting (SFB), our scheme requires lower average bandwidth, and obtains smaller variance. Besides, given a fixed bandwidth smaller than the peak, the SRFS scheme achieves a far smaller blocking rate. The SFDPB scheme also obtains the similar results. Besides, the work investigates another important issue – client-buffer saving. A proposed approach requires a client to delay downloading segments as late as possible. We further prove that a CBR broadcasting scheme with the approach lets its clients buffer the least segments. In addition, the paper presents a reverse segment arrangement, which arranges the segments in descending order of their numeric indexes on each stream. Using the downloading delay and the reverse arrangement, we improve the fast broadcasting (FB) scheme, and obtain the reverse FB (RFB), which requires a client to buffer only 25% of a playing video without sacrificing waiting time. In comparison with the FB, the new pagoda broadcasting, and the recursive frequency-splitting schemes, the RFB saves the client buffers by 50%, 33%, and 33%. Extending the scheme, we propose the RFB-n, which provides a tradeoff between client buffers and server bandwidths. Given three extra streams, the scheme requires a client to buffer only 10% of a playing video. Furthermore, we apply the approaches to other schemes, such as the greedy disk-conserving broadcasting and the pyramid broadcasting, and their buffer requirements can be reduced by 30% to 75%. The article also discusses the limit of the approaches to buffer saving finally. |
