[Babel-users] Some idea on cost calculation based on hop count, loss and RTT

[StarBrilliant]

On Sat, Apr 21, 2018 at 8:32 PM, valent.turkovic at gmail.com
<valent.turkovic at gmail.com> wrote:
> Hi Star,
> what technology do you use for your long range links? Are there 3G
> links with quite low signal quality and therefor you get a high loss
> of packets?
>
> Please explain your setup so we can get a better overall picture of
> what you are trying to accomplish.

Hi Valent,


My setup is a full mesh VPN over UDP running on wired Ethernet. I use
Babel to connect nodes all around the globe.

The architecture includes 2 nodes located in US, 2 nodes located in
Japan, 2 node located in China, and 2 nodes located in Germany. I want
to have maximum TCP throughput between any of them.


I found two problems:

1. A direct link Japan -> Germany takes 380ms RTT at peak hour, while
Japan -> US -> Germany takes 260ms. I want Babel to choose an indirect
link during peak hours.

2. The direct link between two China nodes has 10ms RTT but up to 20%
UDP packet loss at peak hour, because some monopolic ISP punishes
inter-ISP connections. I want packets to detour to Japan then back to
China to reduce packet loss.


The facts are:

1. Babel uses rxcost as 256 by default, so a 120ms difference does not
add too much to the metric. I had to reduce rxcost to 16, but that
leads to the next problem.

2. Link quality estimation scales by rxcost, not the sum of rxcost +
RTT. Thus a high loss would contribute nothing to metric if rxcost is
reduced.