Explore pacing-gain to understand early termination

[gfr10598]

We have a lot of ndt7 data now, and should explore the pacing-gain behavior, to understand how well early termination could work based on pacing-gain, and also verify whether there is an earlier fixed termination time that would be effective for 95+ percent of clients.

[gfr10598]

https://console.cloud.google.com/bigquery?sq=581276032543:5a15dedd9d1a4133b498357ddc54929f

Looks like pacing gain usually drops below 1.25 within less than 5 seconds. Need more analysis.

[gfr10598]

WRONG - bug in SQL

ok1	ok2	ok5	ok10	total
3988220	4233782	4254709	4275755	4275755
--	--	--	--	--

92% converge within 1 second
99% converge within 2 seconds
99.5% converge within 5 seconds
100% converge within 9-13 second test duration.

[gfr10598]

ok1	ok2	ok5	ok10	total
2226638	3189769	3990049	4184933	4249948

Of those that converge:
50% converge with 0.8 seconds
90% converge within 3.3 seconds
95% converge within 4.8 seconds
99% converge within 8.1 seconds

[gfr10598]

The average convergence time is about 1.4 seconds, and average BytesAcked at convergence is about 69MB.
Very fast tests tend to converge faster. Slower tests tend to converge slower. Tests that converge in more than 1 second converge after, on average, about 50MB of transfer.

Spreadsheet (google.com)

[gfr10598]

Found another SQL bug. The spreadsheet has been updated.
The average BytesAcked trends up with convergence time, and averages about 5.5 MBytes. (log mean around 1.7MB)

The worst 5%, with convergence time > 4.8 seconds, average 28 MBytes (log mean 4MB), and the latest converging average around 40 to 60 MBytes.

The BBRInfo.MinRTT averages around 1 msec (log mean 0.88) for the fastest converging tests, and around 200 msec for tests that take 6 seconds or more to converge.

On average, it looks like it takes around 25 to 30 minRTT to converge, but sometimes as few as 5 or 10, and sometimes 1000s of MinRTTs.

[gfr10598]

Median number of round trips to convergence is around 25 for fastest convergence, up to 50 for the slowest convergence.
Query

[gfr10598]

I've refined the query to actually look for first two crossings, from <=1 to <1, and from <1 to >=1. This changes the results slightly, but not dramatically.
NOTE: With this convergence metric, only about 3/4 of the ndt7 tests reach convergence. For another 10%, the PacingGain drops below 1.0, but does not cross 1.0 again.

Spreadsheet (google.com only)
Query

[gfr10598]

BQ connected spreadsheet comparing throughput, BBR_BW at convergence, and BBR_BW at 10 sec.

Convergence vs speed

[gfr10598]

[laiyi-ohlsen]

Needs design discussion with @pboothe