On 8/6/19 7:05 PM, grarpamp wrote:
On 8/6/19, Roger Dingledine arma@torproject.org wrote:
On Tue, Aug 06, 2019 at 05:31:39PM -0400, Rob Jansen wrote:
Today, I started running the speedtest on all relays in the network.
There will be another confusing (confounding) factor, which is that the ... as intended. :) So, call it another thing to keep an eye out for during the experiment.
Someone here posted they were testing with sub-minute durations... tens of seconds. That's unlikely enough to allow TCP to adjust over everything in the circuit to really measure "bandwidth". And is instead likely to be measuring something between "setup latency" and that, with an uncharacterized ramp in the middle.
- That person was Rob, the one who just said they've started their measurements. Rob's original announcement is here[0].
- We've been looking into stuff like this for the last year and have some promising results from sub-minute-duration measurements with measurement hosts spread around the world. This is despite suspected sources of inaccuracy such as TCP slow start and high bandwidth delay products.
- What Rob is doing isn't even trying to get an accurate measurement of a relay's capacity. It's solely to test the hypothesis that observed bandwidth is a poor estimate of capacity*. I refer again to [0] for the motivation, design, etc.
You probably want to be scatterplotting a bunch of different things and durations on metrics.tpo.
And isolating out path nodes and things from whichever it is you're trying to measure. Introducing known inputs. Etc.
Thanks for the input. I'm sure our analysis of the collected data will be ... thorough and exciting.
Matt
* You might argue that an artificial 20 second or even 2 minute burst of traffic is still bad estimate for long term sustained capacity. That may be a good argument. But I'd argue that it's strictly better than the current strategy: keep track of your biggest natural 10 second burst observed in the last 5 days.
[0]: https://lists.torproject.org/pipermail/tor-relays/2019-July/017535.html