| 钟植任* **,潘恒***,武庆华* **,谢高岗** ***.pBBR:面向应用性能偏好的帕累托最优拥塞控制机制[J].高技术通讯(中文),2025,35(7):711~723 |
| pBBR:面向应用性能偏好的帕累托最优拥塞控制机制 |
| pBBR: Pareto-optimal congestion control for application performance preference |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 07. 004 |
| 中文关键词: 拥塞控制;吞吐量-时延;帕累托最优;贝叶斯参数优化 |
| 英文关键词: congestion control,throughput-latency,Pareto optimum,Bayesian parameter optimization |
| 基金项目: |
| 作者 | 单位 | | 钟植任* ** | (* 中国科学院计算技术研究所 北京 100190)
(** 中国科学院大学 北京 100190)
(*** 中国科学院计算机网络信息中心 北京 100190) | | 潘恒*** | | | 武庆华* ** | | | 谢高岗** *** | |
|
| 摘要点击次数: 340 |
| 全文下载次数: 653 |
| 中文摘要: |
| 作为网络传输控制机制的核心,拥塞控制关注如何在异构网络环境中最优化特定传输性能目标。 已有拥塞控制机制忽略了不同应用的性能偏好在吞吐量-时延两个维度上的帕累托最优前沿(Pareto optimal frontier,POF)分布,难以满足差异化应用的性能需求。 针对上述问题,本文提出了一种面向应用性能偏好的帕累托最优拥塞控制机制 pBBR(Pareto-optimal BBR),结合离线网络场景学习和在线控制参数优化的思想,最大程度满足应用的差异化性能偏好。 实验结果表明,pBBR 能够在一个采集-识别周期内判断出网络场景的切换,从而快速选择当前网络场景的最优控制参数。 每个网络场景下,pBBR 都能够最大化满足不同的应用性能偏好:针对吞吐量敏感业务,pBBR 可以达到 Cubic(吞吐优先)的 97% ,且时延只有 Cubic 的 52% ;针对时延敏感业务,pBBR 的时延可以达到 Sprout (时延优先) 的95% ,同时吞吐量损失只有 1% 。 此外,多参数优化可进一步提升 pBBR 性能,例如在高铁长期演进技术(long term evolution, LTE)通信场景下,单参数 pBBR 的吞吐量、时延分别是 Cubic 的 94%和 99% ,而三参数 pBBR 则分别提升到 Cubic 的 101%和 93% (优于 Cubic)。 |
| 英文摘要: |
| As the core of network transmission control mechanism,congestion control focuses on how to optimize specific transmission performance targets in heterogeneous network environments. Existing congestion control mechanisms ignore the Pareto optimal frontier (POF) distribution of performance preferences of different applications in the two dimensions of throughput and delay,making it difficult to meet the performance requirements of differentiated applications. To mitigate the problem above,this paper proposes a Pareto-optimal congestion control mechanism for application performance preferences,Pareto-optimal BBR (pBBR),which combines the ideas of offline network learning and online control parameter optimization to satisfy applications’ preference to the greatest extent. Experimental results show that pBBR can determine the switching of network scenarios within a collection-identification cycle,thereby quickly selecting the optimal parameters for current scenario. In each network scenario,pBBR can maximize the satisfaction of different application performance preferences: for throughput-priority services,pBBR can reach 97% of Cubic (throughput-first),and the latency is only 52% of Cubic; for latency priority services,the latency of pBBR can reach 95% of Sprout (latency-first),while the throughput loss is only 1% . Furthermore,multi-parameter optimization can enhance the performance of pBBR. For example,in high-speed railway LTE scenarios,the throughput and latency of one-parameter pBBR reach 94% and 99% of Cubic, respectively, while the three-parameter pBBR improves these metrics to 101% and 93% (outperforming Cubic). |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
