Analysis of High-speed Data Transfer Protocols

Abstract

Researchers working in diverse fields such as astronomy, experimental physics, genomics, and meteorology frequently have to deal with analysing voluminous amounts of complex data. Such data is often referred to as big data. These researchers work in teams and have to transfer this data over long distances. Many research sites, although equipped with 100 Gb/s links, only use 10 Gb/s links to transfer their data. Efficiently transferring big data over long distances requires the use of appropriate transfer protocols. TCP Reno [62] cannot utilise the high-speed network capacity well while UDP does not provide reliable transmission of data. Sneakernet is also inefficient because of slow disk read/write times. Several TCP-based and UDP-based protocols have been proposed, however, a comparative analysis of such protocols is lacking in the literature. We built in-lab, national and international testbed and used TCP-based and UDP-based high-speed data transfer protocols to measure their performance and fairness. We performed extensive experiments to measure the effectiveness of each protocol in terms of its throughput for various round-trip times, and against increasing levels of congestion inducing TCP or UDP background traffic on a 10 Gb/s network. From the results, we have analysed the protocols to choose a protocol that is well-suited for any particular network path. Also, there are definitions of ‘fairness’ for traditional network flows but these do not consider the flows’ purpose. We investigate how different high-speed data transfer systems perform on high-speed communication networks and observe congestion induced by the systems on other traffic in the network. Measuring induced congestion may provide us insights of what is ‘fair’ in high-speed networks, and how it is different to traditional fairness where the purpose of a flow is not considered. These insights could help both users and network administrators to consider network flow fairness in a new light. We found the primary bottleneck of the high-speed data transfer is in slow disk speed and small TCP socket buffer size. Using multiple transfer flows improves the performance of GridFTP and FDT in a long-haul link more than in a short-haul link because it increases aggregated TCP socket buffer size, and has less impact on the shared link in terms of increased RTT. However, using multiple flows causes more packet losses on the path and each flow goes to recovery mode which may take longer than short-haul link. We found that GridFTP is the most efficient protocol for transferring data in high-speed networks. We recommend using fewer than four parallel flows with GridFTP because of the potential impact to other network services due to increased RTT on shared links. This will improve the existing data transfer performance and reduce its impact on shared links.