Authors:
Mariam Kiran - ORNL
Cong Wang - RENCI
Imtiaz Mahmud - LBL
George Papadimitriou - ISI
Date:
July 14th, 2023
Title:
Studying TCP Fairness in Large Transfers over High Throughput Links
Abstract:
As network resources are limited, there is an increasing need to understand how TCP flows share a network link and how their performance is impacted. To the best of our knowledge, there is limited work on how they influence each other, and most current work relies on simulation results at comparatively low BW settings, not reflecting real-world network challenges. In this paper, we present novel findings on TCP fairness by comparing TCP Cubic, Reno, Hamilton, and BBR versions 1 and 2 across multiple real-network sites with high-speed BW supporting up to 25 Gbps. Our study examines the TCP behavior when routers implement FIFO, FQ\_CODEL, and RED as AQM algorithms, along with various buffer sizes. Our results demonstrate that varying buffer sizes and queuing techniques can yield vastly different outcomes for different network BWs, with BBRv2 emerging as one of the best fair algorithms to use. These findings hold a significant importance for high-speed network transfers, particularly in large scientific data transfers. Additionally, we provide comprehensive analysis and insights on how network fairness can be achieved for future network design.
Негізгі бет 20230714 - ORNL, RENCI, LBL, & ISI - “Studying TCP Fairness in Transfers over High Throughput Links”
Пікірлер