Should paced TCP Reno replace CUBIC in Linux?

摘要

High-speed protocols such as CUBIC and Compound TCP have been proposed for fast data transfer over long-distance networks. Currently, CUBIC and Compound TCP are the default transport protocols in the Linux and Windows operating system, respectively. However, both CUBIC and Compound TCP are known to have performance issues. Packet pacing can significantly improve performance of TCP Reno in high-speed networks. To that end, we examine the performance of paced Reno with CUBIC, and Compound TCP over a single bottleneck link with Drop-Tail buffers. Using extensive simulations, we show that in a many flows regime, with or without selective acknowledgements (SACK) support, paced Reno performs at the most, or better than CUBIC and Compound TCP. We also show that, with large Drop-Tail buffers, flow synchronization effects give rise to non-linear oscillations in the queue size dynamics, leading to low link utilization, aggregate throughput and fairness. On the contrary, small buffers ensure stable queues, and help paced Reno to achieve better network performance. With such incentives, paced Reno can possibly replace CUBIC in the Linux OS.