The end-to-end performance effects of parallel TCP sockets on a lossy wide-area network

Abstract

IntroductionThere are considerable efforts within the Grid and high performance computing communities to improve end-to-end network performance for applications that require substantial amounts of network bandwidth. The Atlas project [19], for example, must be able to reliably transfer over 2 Petabytes of data per year over transatlantic networks between Europe and the United States.Recent experience [1,2] has demonstrated that actual aggregate TCP throughput realized by high performance applications is persistently much less than the end-to-end structural and load characteristics a network indicates is available. One source of poor TCP throughput is a packet loss rate that is much greater than what would be reasonably expected [20]. Packet loss is interpreted by TCP as an indication of network congestion between a sender and receiver. However, packet loss may be due to factors other than network congestion, such as intermittent hardware faults [4].Current efforts to improve end-to-end performance take advantage of the empirically discovered mechanism of striping data transfers across a set of parallel TCP connections to substantially increase TCP throughput. As a result, application developers and network engineers must have a sound understanding of how parallel TCP connections improve aggregate throughput as well as their effects on a network. This paper addresses several questions concerning the use of parallel TCP connections. The first question is how the use of parallel TCP connections increases aggregate throughput. The second is how to determine the number of TCP connections needed to maximize throughput while avoiding network congestion. Finally, understanding how parallel TCP connections affect a network, and under what conditions they should not be used. This paper suggests some practical guidelines for the use of parallel sockets to maximize end-to-end performance for applications while simultaneously minimizing their network effects.The remainder of this paper is organized as follows. Section two discusses current work. Section three presents a parallel socket TCP bandwidth estimation model and the experimental results. Section four discusses the behavior of packet loss on the Internet and its effect on TCP throughput. Section five presents conclusions and guidelines for using parallel sockets, and discusses some possible avenues for future work. Current WorkApplications generally take two approaches to improve end-to-end network throughput that effectively defeats the congestion avoidance behavior of TCP. The first approach utilizes UDP, which puts responsibility for both error recovery and congestion control completely in the hands of the application. The second approach opens parallel TCP network connections and "stripes" the data (in a manner similar to RAID) across a parallel set of sockets. These two approaches are aggressive and do not permit the fair sharing of the network bandwidth available to applications [5].Recent work [1,2,6] has demonstrated that the parallel socket approach ...

A simulation-based study of HighSpeed TCP and its deployment

“…When competing with connections over a congested link, each of the parallel streams will be less likely to be selected for having their packets dropped, and therefore the aggregate amount of potential bandwidth which must go through premature congestion avoidance or slow start is reduced. An application opening N multiple TCP connections is in essence creating a large virtual MSS on the aggregate connection that is N times the MSS of a single connection [25].…”

“…We used the simulator error model to simulate losses in the bottleneck link. This loss model was set to drop a The theoretical performance of parallel TCP streams over a range of packet loss rates follows the equation presented in [25], for the condition of this experiment (MSS = 1500 bytes, RTT= 100 ms, C = 1, and packet losses impacts parallel streams to the same extent). For the response function of HSTCP was used the equation defined in [18].…”

Orchestrating parallel TCP connections: Cyclic and probabilistic polling policies

“…The model we study in this paper describes situations in which several TCP connections are opened by one user between the same source and destination. Such models have been extensively studied in the literature (see e.g., [5,10,11]). Losses (which is interpreted as a congestion signal) are used as signals to reduce the window size of one of the connections.…”

Improving throughput and maintaining fairness using parallel TCP

“…Parallel TCP streams are aggressive on a shared network and can steal bandwidth from competing TCP streams. Previous work [22] showed that a single application using N parallel TCP streams competing with k other streams will receive N/(N+k) of the network bandwidth, rather than the 1/(N+k) portion the other streams receive.…”

“…In previous work [22], we performed a series of network measurements between the University of Michigan in Ann Arbor and NASA AMES in Moffett Field, CA over a high speed path to determine if parallel TCP streams were effective at improving throughput, and to gain insight on the relationship between the number of parallel TCP streams and performance. We found that as the number of TCP sockets used in the parallel stream increased, aggregate throughput increased linearly until the bandwidth * delay product of the network was achieved.…”