針對雲端與產業化資料中心網路研發設計一套更快速且更可靠的網際網路通訊機制

Abstract

本計畫的目標是針對產業化資料中心網路設計一套低延遲、高throughput的傳輸機制。目前，產業化資料中心網路仍以TCP作為主要的傳輸協定，但是已有許多研究指出TCP在資料中心網路下運作會產生三種問題。首先，用來建構產業化資料中心網路的網路交換機具有shallow-buffered的硬體特性，使得封包遺失問題更加嚴重。其次，因為應用程式具有partition/aggregate之特殊程式結構，導致大量回應訊息的同步問題。最後，因為資料中心網路內的statistical multiplexing程度較低，若使用傳統TCP調整congestion window的方式將會犧牲throughput。為了改善上述問題，我們提出了Bandwidth-Aware TCP (BATCP)。一個能夠主動探測網路的可用頻寬，快速適應頻寬的動態改變，避免對交換機造成週期性的緩衝區溢位，並且根據網路的狀況調整探測行為的壅塞控制機制。相較於傳統的TCP，BATCP能夠以更有效率的方式利用交換機上有限的緩衝區。實驗數據顯示BATCP具有較好的回應時間、較低的封包遺失率、以及較高的throughput。

This work aims to develop a low latency and high throughput transport mechanism for commoditized data center networks. Within the data center network, TCP is the dominant protocol for data transferring. However, previous researches indicated three inefficiencies of TCP under the data center environment. First, because commoditized switches of which data center network is composed are shallow-buffered, packet loss becomes more severe. Second, due to the partition/aggregation structure of applications, massive response messages are synchronized. Third, the degree of statistical multiplexing is very low in data center network, the reaction of traditional TCP to congestion events, cut the congestion window in half, leads to loss of throughput. To alleviate the above inefficiencies, we proposed a Bandwidth-Aware TCP (BATCP). BATCP is designed to achieve active probing the available bandwidth, adapting its sending rate to dynamic changes in bandwidth rapidly, preventing switches from buffer overflows periodically, and adjusting the probing behavior according to the network condition. In comparison of traditional TCP, BATCP utilizes switches’ buffer in a more efficient way. Experimental results showed that BATCP can achieve better response time, lower packet loss rate, and higher throughput.