(The) adaptive scheme for adjusting the aggregate throughput of parallel TCP flows

Abstract

The Internet bandwidth has exponentially increased over the past few decades. As the Internet bandwidth grows, new data-intensive applications such as climate simula- tion, earth observing, and high energy physics are emerging. Terabyte and petabyte data need to be transferred over the Internet. Many of these applications use the Transmis- sion Control Protocol (TCP) for the reliable data transfer. Since the standard TCP, TCP Reno, was developed in 1988 [1], it has been widely adopted and performed remarkably well. However the standard TCP eventually underutilize high bandwidth networks [2]. To solve the low utilization problem of the standard TCP, many new TCP protocols such as HSTCP [2], STCP [15], FAST [15], BIC [17], and CUBIC [18] emerged. Even those studies improve the low utilization problem of the standard TCP, they are concerned in only a single TCP flow. In this paper, we research into the use of parallel TCP flows (we call it the parallel TCP) on high bandwidth and delay networks. The parallel TCP has been an easy way to achieve higher throughput. Researchers have already developed tools such as GridFTP [6] , bbcp [7], and pSocket [8] which open multiple TCP flows in parallel. Every paral- lel TCP flow increases its own congestion window size concurrently per RTT, and only some of them, which experience a packet loss, decrease their own congestion window size by half. Therefore the aggregate throughput of the parallel TCP increases quickly and it 5 is less sensitive to a packet loss (or congestion) than a single TCP flow (we call it the single TCP). It is why the parallel TCP can achieve the higher aggregate throughput as compared to the single TCP. The parallel TCP can be on not only a single machine but multiple machines. Over high bandwidth and delay networks, a bottleneck can be at an end system due to per- formance limitation at disk, bus, and/or network interface card in a single machine [3]. Hence, the parallel TCP a...