[reno] better handling of high-BDP connections

[kazuho]

Adds a flag called highbdp_mode that changes the increase ratio.

The increase ratio of ordinary Reno is 1 mtu per round-trip (i.e. CWND being acked).

When in high-BDP mode, the CC uses exponential increase during congestion avoidance phase, where the increase rate is calculated so that it would reach the send rate at which the loss was observed every ~1 second. This is roughly equal to what Cubic does.

[kazuho]

Below are the loss observations when sending at fullspeed from Tokyo to West Coast for ~15 seconds:

The first column is current time in milliseconds.

Ordinary Reno:

**** 1624356443776 - reno_on_lost:74 cwnd=2211062,rtt=279
**** 1624356444282 - reno_on_lost:74 cwnd=1547743,rtt=313
**** 1624356447150 - reno_on_lost:74 cwnd=1090588,rtt=318
**** 1624356451871 - reno_on_lost:74 cwnd=779539,rtt=224

High-BDP Reno:

**** 1624356207664 - reno_on_lost:98 cwnd=2182804,rtt=289
**** 1624356208168 - reno_on_lost:98 cwnd=1603280,rtt=302
**** 1624356208699 - reno_on_lost:98 cwnd=1208808,rtt=311
**** 1624356210035 - reno_on_lost:98 cwnd=1232443,rtt=307
**** 1624356211309 - reno_on_lost:98 cwnd=1226988,rtt=304
**** 1624356212604 - reno_on_lost:98 cwnd=1239536,rtt=307
**** 1624356213914 - reno_on_lost:98 cwnd=1246266,rtt=309
**** 1624356215167 - reno_on_lost:98 cwnd=1229620,rtt=308
**** 1624356216480 - reno_on_lost:98 cwnd=1240246,rtt=308
**** 1624356217719 - reno_on_lost:98 cwnd=1228286,rtt=307
**** 1624356219007 - reno_on_lost:98 cwnd=1230350,rtt=307
**** 1624356220315 - reno_on_lost:98 cwnd=1239375,rtt=309
**** 1624356221605 - reno_on_lost:98 cwnd=1239595,rtt=310
**** 1624356222882 - reno_on_lost:98 cwnd=1233795,rtt=308
**** 1624356224172 - reno_on_lost:98 cwnd=1238141,rtt=309

As can be seen, ordinary Reno does not have chance to recover from loss. Compared to that, high-BDP reno is testing the peek every ~1 second.

[kazuho]

Use of exponential increase during congestion avoidance is not expected to cause difference from ordinary Reno when using beta of 0.7. The chart below is the buffer under-utilization curve of Cubic (blue), Reno (purple), high-BDP Reno (brown). One loss interval of ordinary Reno is normalized to 1 second.

When there is increase in available bandwidth, use of exponential increase provides benefit if not as good as Cubic. The chart below shows the buffer under-utilization curves when available bandwidths becomes 2x at the moment of congestion. In case of Cubic, it catches up with the new available bandwidth at ~2.2 seconds (notice when the curve reaches plateau). In case of ordinary Reno with loss interval normalized to 1 second, it takes ~3.3 seconds. With high-BDP Reno, ~2.5 seconds.

[kazuho]

What if there's a large ack that would have result in count > 1, but calc_highbdp_inrrease return something like max_udp_payload_size? Current code adopts the latter, while it should adopt the former. (credit to @janaiyengar).

[janaiyengar]

With the fix, this looks great, thanks @kazuho !

[kazuho]

Closing in favor of #470.