The last sermon on congestion control-關於擁塞控制研究的勸退文

 In my opinion, BBR [1] is the last solution in congestion control area. Even the origin version of BBR also brings out some issues e.g, favor towards on long RTT flows and bandwidth allocation unfairness in links with shallow buffer, there is solutions (BBRPlus [2])to remedy these problems. The bandwidth allocation can be guantateed even in shallow buffer links and lower tranmission delay can be achieved in BBRPlus. There is a reveiw paper [3] for interested readers, in which the performances of BBR, BBRPlus and BBR v2 are evaluated on ns3. The conclusion is that aplliying BBRPlus is highly recommended.
 To design new congestion control algorithm outperforms BBR algorithms is quite hard. There is a trend to apply reinforce learning on congestion control[4,5,6].
The author of Remy [4] conducts experiments on ns2. It seems only two flows are trained. No tests of Remycan be available on real networks. The code of Remy was relesed on github. But the author claimed that Remy is not ready for real internet [7].
 正如題目的副標題，這是一篇擁塞控制研究的勸退文。但是呢，準備論文造假的，則不在勸退之列。
 從1989年算起，擁塞控制的歷史將近30年了。傳統的擁塞控制算法主要思想，根據網絡中的信號判斷網絡是否處在擁塞狀態；若鏈路處在擁塞狀態，發送端降低數據包的發送速率，緩解網絡擁塞。But these signals that can be used to determine whether bottleneck link falls into congestion are limited, including packet loss, round trip delay, one way delay and delay gradient. 信號的有限，就限制了可以玩出的姿勢。在BBR之前，大部分算法，都是在AIMD算法之上修修補補。而且，很多論文根本沒有給出代碼，效果到底如何，值得懷疑。究竟有沒有做出來，都很值得懷疑！
 兩年來，我寫了一些關於擁塞控制博客以及仿真分析。期間，有些碩士生向我請教一些關於擁塞控制的問題。但是我最後都要奉勸，換個容易水論文的方向吧，擁塞控制已經沒什麼可做的了。不知道是什麼心理作祟，大部分人十分固執，毅然決然地跳進這個坑。到臨近畢業，也沒有從這個坑裏爬出來。這些同學的心理，我猜測，無外乎：要麼覺得自己特別，只要努力，總是能夠做出一些成果；要麼覺得同行是冤家，覺得我的勸退是爲了消除競爭。我青春的時候，也覺得自己定能做出一番事情。那種心理，如古時候的猶太人，覺得族人是天選之人–destined to the promised land flowing with honey and milk. 隨着年齡的增長，我越發認識到自己的普通。對於很多人來說，這輩子最special的時刻，可能就是形而爲人的時刻。
 有些研究方向，即使努力了，也未必能夠做出結果，擁塞控制就屬於這種。大概是2017年10月份，我從BBR開始閱讀一些關於擁塞控制方向的論文。這兩年，我基本沒有在擁塞控制領域發出什麼文章，只是寫了兩篇review[3,8]。
 我現在是讀博第五年，決定不要搞這個了。
[1] BBR: Congestion-Based Congestion Control
[2] BBR Plus
[3] An Evaluation of BBR and its variants
[4] TCP ex Machina: Computer-Generated Congestion Control
[5] QTCP: Adaptive Congestion Control with Reinforcement Learning
[6] TCP-Drinc: Smart Congestion Control Based on Deep Reinforcement Learning
[7] Remy is not ready for real internet
[8] Congestion Control for RTP Media: a Comparison on Simulated Environment
[9] 來自Google的TCP BBR擁塞控制算法解析