BBR encounters cubic

BBR Dynamics is only in pure BBR Match the model , Otherwise, it requires various skills to coexist with other algorithms ,BBRv2 This is the compromise .
This article focuses on BBR And ProbeRTT.

This state is interesting , Its convergence mechanism is so exquisite . Take one big one small two streams as an example , Big flow means it invades buffer, Enough to affect minrtt collection ：

• The stream enters ProbeRTT, Qingrang buffer, Large and small streams are collected at the same time minrtt.
• Small streams enter ProbeRTT, Too large collected minrtt, Wait for the big stream to enter ProbeRTT, Two streams are collected at the same time minrtt.

As long as all BBR The flow configuration is the same ProbeRTT window , The convergence point is naturally synchronized , Namely BBR bandwidth / Time Those small tips protruding downward on the picture .

Like those joints on elevated bridge decks or railway tracks , At a fixed distance, you have to click ,BBR This problem also exists for some streaming media , Synchronous Caton . randomization ProbeBRR Window can solve , such as 5s～15s Enter randomly in ProbeRTT.

Each stream ProbeRTT Interval random , It does not affect that all streams are still collected at the same time minrtt, This is the ultimate .

All streams are collected minrtt From that moment on ,BBR stay ProbeBW The state converges to fairness , See below “ProbeBW state probe Stage BBR How to converge fairly ” section ：
The math behind dynamics of TCP BBR

Go out of the ideal , Come to the real world .

BBR No peace CUBIC Dynamics of fair convergence , Only by adjusting parameters .

Found in practice and CUBIC coexistence , No robbing buffer,BBR I can't stand firm , Unable to stably maintain the state machine , At this time, the bandwidth data is still higher than CUBIC good , Not benefit from BBR Model of , It is ：

• BBR Can hold 10 rounds Of maxbw Don't give in .
• BBR The packet loss state still tries to stabilize the state machine and after recovery undo.

But the retransmission rate has also increased , This has been overdrawing resources to maintain the state machine . as for ProbeBW Medium probe/drain, Almost completely disordered , because BDP No longer credible .

And CUBIC coexistence ,ProbeRTT Will give BBR Bring bad luck instead of opportunity , Suddenly take the initiative to lower inflight Gave other streams buffer The illusion that is still not full , etc. BBR sign out ProbeRTT when , Still unable to collect minrtt, On the contrary, the transferred bandwidth is also occupied .

ProbeRTT The goal is to collect minrtt, When it is confirmed that the goal cannot be achieved , No more ProbeRTT It's a good plan , Rather than 4 inflight Change to 70%, The latter is still unavailable minrtt, Why .

It seems that there are two choices ：

• Reduce minrtt The expectations of the , maintain CUBIC State, midrtt As minrtt.
• according to BBRv2 How to do it , Proportion reduction inflight.

Ideal BBR The posture of is easy to be realistic buffer Upset , It depends too much on pacing , Or above , see “probe Queuing problem caused by crowding and its solution ” .

Yes trick Can cope with , Long flow is effective .

When entering ProbeRTT Measure before and after srtt and delivery rate, See if the result has changed greatly , Distinguishable RTT Raise whether you do it yourself , If it weren't for what I did , Then it is recorded in the link portrait .

If oneself ProbeRTT,RTT Still high , You can judge that you haven't occupied buffer, If similar data is collected in a window minrtt, Then the other party is BBR, Don't go in ProbeRTT The probability is reduced , Otherwise increase . Every time I enter ProbeRTT Whether or not , It all depends on this probability , This will ProbeRTT And collection minrtt Separate , No longer dependent on .

Explain it. ：

for the first time ProbeRTT after , At least you can guarantee yourself inflight yes 4( perhaps 10), The increment at this time RTT It must be someone else , The next time ProbeRTT, if RTT Still high , The probability can be explained by others , Because the BBR Rules don't occupy buffer, Since it's not what I did , Again inflight Can't collect minrtt, Pure loss , I will not enter this time ProbeRTT.
minrtt You don't have to pass every time inflight = 4 get , confirm buffer It's not what I did , Can inherit recently relatively trusted minrtt, Or take twice 10s Room minimum RTT, Take into account the queue delay caused by others .

All in all , We should continue to depict a link portrait in the connection , It will become more and more accurate .

CUBIC, BBR,25 Gbps,50 us link , No packet loss ,1000 pkts buffer,5v5, Look at the effect ：

Parameter adjustment probability 10 Double gain , The original intention is BBR stay 150 Hang it in seconds CUBIC. Modifying preferences can maintain fairness , But it doesn't make a difference if you don't hang it .

Ben is going to Nanhui to catch the sea and listen to the tide , Although I can't compete with Meisha and Xiaomeisha in Yantian, Shenzhen , Dapeng Nan'ao bayberry pit ratio , But at least it's also a sea … But Xiaoxiao has an exam tomorrow , You can't go , I wanted to debug this in the hotel tonight , In this hot afternoon, it's over . Don't release the code this time , Because you can't always put , even stap poc Don't let go . Last night, a friend suggested , Compared with those authors who post a official account QR code behind their long speeches , And those big factories P comparison , I'm just SB, They took too much output from me , But I still seem happy . My aftertaste is endless , Thank you very much for your kindness , I feel happy , But I'm always misunderstood about what benefits I get , But in reality, I have to disappoint my wife and children even for a trip to Jiangsu, Zhejiang and Shanghai because I have no money . But anyway , In my opinion, communication is the purpose , Anyway, I haven't been in this business for a long time , At present, we don't rely on this to eat , Or write some ideas , There is no conflict with anyone ,SB Just SB Well .

Zhejiang Wenzhou leather shoes wet , It's not fat when it's raining .