BBR and queuing

BBR Why can't it be cashed no queuing Commitment ？

First, clarify a fact ：

• Many stream mixing parts of different algorithms are used in modern high bandwidth Heterogeneous Networks , Single stream behavior can hardly change the statistical characteristics of multi stream multiplexing bandwidth .

BBR Suddenly collected small bandwidth, it 10 rounds Of max-filter The wheel is big enough , It's easy to cross this pit ,BBR The shock absorption effect for concave surface is very good . but BBR There is a problem , If you suddenly collect large bandwidth, It remembers , Even a noise , Also maintain 10 rounds, During that period BBR Will continue overestimate bandwidth, So as to continue queuing.

Multi algorithm mixed part , Scenario of multi stream multiplexing bandwidth , Suddenly collected big bandwidth The probability is very high , At least with the sudden acquisition of small bandwidth The probability is the same , For example, an elephant stream suddenly exits , This stream will sell a lot of bandwidth , But these bandwidths will be instantly flooded by multiple streams , A sudden short burst will consume a lot in an instant buffer equally .

No matter concave noise , Or convex noise , Should be smoothed .BBR Have the ability to smooth concave noise , But there is no ability to smooth convex noise . therefore BBR In the scenario of multi algorithm and multi stream statistical multiplexing , Will continue to be counterintuitive queuing, and BBR paper The description is inconsistent .

stay BBR Single stream or BBR Multi stream scenario ,paper Description of the BBR Dynamics can ensure the final convergence point no queuing, But multi algorithm and multi stream scenarios , BBR Instead of crossing, you will remember convex noise ,overestimate bandwidth Will stack .

Such as noise A By max-filter remember ,BBR Will be with overestimate Of A As a benchmark pacing rate, If noise is collected again A + d, be A + d Will become the new benchmark pacing rate, At least in the 10 rounds Effective within , And so on , therefore BBR Of overestimate It is stackable . It's very helpless .

What do I do ？ Two clear points ：

• With acquisition bandwidth The average shift of fingers as max-filter The input of , Or use max-filter Output of and current acquisition bandwidth Do shift average . Finger shift averaging smoothes noise ,probe Although the bandwidth is slow , But steady .
• ProbeBW State, , No probe Bandwidth to , Even if it's bigger , Do not enter max-filter, Only up probe Get extra bandwidth before doing maxbw to update . External noise input can easily disrupt the internal state machine decision .

See again ,BBR Also a paper Algorithm , It just happens to perform better in most actual scenarios CUBIC good , It's good . All for BBR Second optimization of , All of them are paper Algorithm , No one can escape Why We Don’t Know How To Simulate The Internet The curse of .

Don't expect to adjust a flow cwnd perhaps pacing rate, It can affect the congestion state of the whole link , This is not the stream you adjust to compete with another stream , But hundreds of streams are playing games , The overall behavior of all flows can only produce a statistical distribution effect .

Of course , And don't make assumptions about any behavior of any flow that implements any algorithm , Whether the behavior is reasonable , Still unreasonable , Is expected , Or unexpected . All strategies should be based on statistical characteristics , Not the assumption of the algorithm .

The current network bandwidth is several orders of magnitude higher than that of more than a decade ago and earlier , The understanding of congestion control needs to be changed as follows ：

• It is difficult for a single stream to fill the bandwidth , Most traffic is app-limited or rwnd-limited, No longer cwnd-limited.
• Delay jitter is hardly caused by single stream , The adjustment of single stream congestion control strategy has little impact on the average distribution of the overall statistical data .
• The benefits of aggressively sending bandwidth runs are getting smaller .
• High bandwidth may be bufferbloat The high delay and bad bandwidth produced by the run , It seems that inflation can't get money , Who started more people , Get more money , But it's also costly .

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