An Efficient Error Prediction and Recovery for Multicast Traffic in Data Center Networks

In cloud application services, data are transmitted through multicasts in scientific computation and online searches to avoid transmission of duplicate packets and reduce bandwidth waste. However, the data amount in data center networks (DCNs) has increased to the scale of big data over time. Big data can be processed only through DCNs by using distributed computing such as MapReduce. Consequently, a large flow of multicasts has been generated. DCNs comprise a high number of commodity servers and switches, which exhibit limited capabilities to process packets. Subsequently, the rate of multicast congestion in DCNs increases substantially, resulting in severe packet loss and transmission error. Packet loss and transmission error cause all transmitting devices to retransmit data, resulting in network bandwidth waste, transmission delay, and disrupted transmission of other data. Consequently, network congestion is exacerbated, and the entire multicast tree crashes. Therefore, this thesis proposes an error prediction and recovery multicast, which enables areas of errors to be effectively predicted and reduces the chain effects of errors by using repair multicast trees and multiple multicast trees that share error information, thereby effectively reducing the costs of reconstructing multicast trees. Simulation results show that the proposed scheme significantly improves the costs of reconstructing multicast trees and the transmission delay.在雲端應用服務中，不論是科學計算或是網頁搜尋，都會使用群播的方式傳送資料，避免傳送重複性封包，減少頻寬浪費。然而，現今在資料中心網路上的資料量與時俱增至巨量資料的規模，這些巨量資料的處理僅能透過資料中心，利用如MapReduce般的分散式計算去得到運算結果，因此相對也產生龐大的群播流量。 由於資料中心是由大量的平價伺服器和交換器組成，這些交換器處理封包的能力有限，這使得群播流量在資料中心網路發生壅塞導致封包遺失或傳送錯誤的機會大幅增加，封包遺失或傳送錯誤的連鎖效也更為嚴重。只要發生封包錯誤或遺失，所有傳送端就必須再重傳資料，這樣不僅浪費網路頻寬，延誤傳送時間，甚至干擾其他資料傳送，使得網路的壅塞更嚴重，進而導致整個群播樹崩潰。 因此，本論文提出有效的錯誤改善機制(Error Prediction and Recovery Multicast ，EPRM)，此方法不僅可以有效地預測錯誤的修復範圍，也可以利用群播修復樹和多個群播樹共享錯誤資訊去降低錯誤發生的連鎖效應，並有效減輕傳送巨量資料的群播樹之重建成本。誌謝 i 摘要 ii Abstract iii 目次 iv 圖目次 v 第一章 緒論 1 1.1 簡介 1 1.2 研究動機 5 1.3 論文架構 8 第二章 背景知識 9 2.1 前言 9 2.2 網際網路的可靠性群播 11 2.3 DCN中的可靠性群播 12 第三章 研究方法 13 3.1 設計構想 13 3.2 設計方法 16 3.2.1 Error Forecast Mechanism 17 3.2.2 Repair Multicast Tree 20 3.2.3 Error Information Sharing 22 第四章 數學分析 23 4.1 分析情境 23 4.2 The Cost of EPRM in Fat Tree 25 4.2.1 The Minimal Cost of Repaired Multicast Tree in Fat Tree 27 4.2.2 The Cost of EPRM and RDCM in Fat Tree 34 4.3 The Cost of EPRM in BCube 36 4.3.1 The Minimal Cost of Repaired Multicast Tree in Bcube 39 4.3.2 The Cost of EPRM and RDCM in Bcube 44 4.4 The Validation of EPRM 46 第五章 實驗結果 49 5.1 實驗環境參數 49 5.2 Repairing Capability of EPRM 52 5.3 Cost'Retransmission and Cost'Delay Evaluation 54 第六章 結論 57 參考文獻 5